1. 具有单粒子灵敏度的聚苯乙烯纳米球的 SRS 成像

众所周知，SRS 显微镜比常规拉曼成像快几个数量级（25、26）。SRS显微镜的成像速度显著提高，因此在粒子成像方面具有高吞吐量。然而，高速 SRS 是否比常规拉曼具有更好的检测限，以及它是否真的能达到纳米塑料的单粒子灵敏度，尚不明确。理论量化有助于首先解决这个问题。对于给定的主要类型的塑料聚合物，我们可以根据塑料密度估算直径为 100 纳米的纳米塑料的质量，并通过其分子量计算重复单元（即构成单体）的数量。如SI 附录表 S1所示，对于大多数主要塑料类型，这个数字约为 106 ^，基于此我们可以进一步估计单个塑料颗粒中最丰富的化学键的数量约为¹⁰⁷。

这样，我们就可以从理论上解释为什么100nm的纳米塑料颗粒难以被传统的拉曼显微镜检测到。典型的C–H振动的自发拉曼截面约为10−29cm2 。^因此， 100nm纳米颗粒的自发拉曼截面为10−22cm2 。在高数值孔径显微物镜下^，激光腰区可以缩小到约2×10−9cm2 ^{。那么每个}激发光子发生拉曼散射事件的概率为(10−22cm2 ⁾ /(2× ^{10−9cm2 ) =} 5 ^× 10−14 。假设使用传统 532 nm 激光器，激光功率较高（10 mW），激发通量为 3 × 10 ¹⁶光子/秒，采集时间较长（100 ms，一张 128 × 128 的小图像需要半小时），那么通过自发拉曼散射，每个粒子总共只能产生约 130 个光子。考虑到整个仪器（包括物镜、滤光片、针孔、光谱仪和相机）的量子产率通常约为 1%，最终只能检测到大约 1.3 个光子。如此微弱的信号很容易被其他背景噪声（例如自发荧光）所掩盖。

通过使用额外的相干斯托克斯激光器，SRS 通过量子刺激放大特定光谱模式（由泵浦和斯托克斯激光器之间的能量差定义）的微弱散射截面。当使用脉冲窄带斯托克斯激光器（24、33）时，受激拉曼增强因子可最大化至¹⁰⁸以上（32、34）。每个泵浦激发光子发生受激拉曼散射事件的概率变为 5×10−6 ^，该概率以泵浦光束针对 C–H 振动所经历的受激拉曼损失来测量。在高速 SRS 显微镜采集（18µs/像素）下测得的泵浦光束噪声为 5×10−7 ^（图1），比单个 100nm 塑料颗粒预期的受激拉曼损失信号低约 10 倍。因此，我们预测窄带受激拉曼光谱将打破自发拉曼光谱的可检测性障碍，并在短短几十微秒内将单个纳米塑料颗粒带入检测范围。

然后，我们使用标准塑料颗粒通过实验验证了检测灵敏度。聚苯乙烯是日常生活中广泛使用的常见塑料之一。特定尺寸的聚苯乙烯颗粒可作为分析标准在市场上买到，并且通常用作研究微纳米塑料的模型材料（35、36）。聚苯乙烯的拉曼光谱表明，苯环上芳香族 C–H 振动在 3,050 cm ^{−1处有一个突出的峰（} SI 附录，图 S1），通过调整泵浦和斯托克斯光束的差异以匹配该跃迁能量，可以选择性地放大该峰以进行 SRS 成像。使用从 100 nm 到 3 µm 的商用 PS 微纳米球，我们评估了我们的 SRS 显微镜在成像纳米塑料方面的检测灵敏度。为了在成像过程中稳定颗粒，我们将稀释的 PS 颗粒嵌入琼脂糖凝胶中。随着粒径变小，在 3,000 cm ⁻¹左右的水背景残留开始占据主导地位（SI 附录，图 S2 a），掩盖了单个 PS 纳米颗粒的真实光谱。为了解决这一背景问题以获得更好的成像对比度，我们用 D _{2 O 代替普通 H 2} O 来制备琼脂糖凝胶（SI 附录，图 S2 b _{）。与 H 2} O相比，D ₂ O的拉曼光谱红移至静默区域（2,200 至 2,800 cm ⁻¹，SI 附录，图 S3），为探测 C-H 振动创造了无背景环境。

因此，可以通过高通量单通道窄带成像测量单个颗粒的SRS强度（2秒内在一个51×51μm FOV中测量约1,000个颗粒，SI附录，图S4） 。该成像速度比其他纳米塑料成像技术（例如AFM-IR和STXM（21、23、37））快几个数量级。在光学衍射极限下，SRS显微镜的最佳空间分辨率为365nm（图1H和I ）。通过200nm像素大小的空间采样进行高通量成像，可以从图像中辨别出500nm以上的单个PS纳米球及其形状（图1D - G）。当颗粒尺寸小于衍射极限时（图1A - C ），有限的光学分辨率会使颗粒图像呈现衍射极限图案。然而，根据衍射极限图和强度分布（SI附录，图S5），单个粒子的SRS强度在100纳米以下仍可轻松识别。因此，我们通过实验证明，与常规自发拉曼光谱相比，SRS成像可以为纳米塑料分析提供更高数量级的成像速度/吞吐量和更高的检测限。

SRS 信号的对数与  ) 以及小于 0.7 µm 的 PS 颗粒直径的对数（图 1 J和SI 附录，补充说明3）。范围内斜率为 2.98 的趋势线表示 SRS 信号（  ) 随着颗粒体积线性增加，而颗粒体积随着颗粒直径的增加而立方增加。当颗粒尺寸依次在 x、y 和 z 维度上放大到填充有效焦点体积时（SI 附录，图 S14），线性依赖性消失。这种良好的线性（ R ² = 0.998）归因于 SRS 信号对目标分析物浓度的基本线性依赖性，在几个方面提供了强大的实用性。首先，可以根据获得的校准曲线估算衍射极限以下颗粒的实际尺寸（SI 附录，图 S16 a），扩展尺寸表征极限。其次，利用已知的塑料密度信息，相同的校准曲线可以转化为参考，以从检测到的 SRS 纳米塑料图像中推断出颗粒质量（SI 附录，补充说明3和图 S16 b）。最后，以 SNR 为 1 作为阈值，可以确定我们的窄带 SRS 显微镜的检测极限（图 1K ），以达到小至 60 纳米的 PS 纳米球。

2. 利用高光谱SRS成像技术对纳米塑料进行化学鉴定的基本挑战

纳米灵敏度解决了首要问题，确保塑料颗粒可检测。技术的化学特异性对于区分塑料与其他共存物质，并进一步区分不同的塑料聚合物也至关重要。受激拉曼散射（SRS）显微镜利用振动光谱作为成像衬度，原则上具备化学成像所需的特异性。在仪器方面，我们利用光谱聚焦技术进行高光谱受激拉曼散射（SRS）成像（38 , 39）。为了在仪器的调谐范围（790 至 910 nm）内最好地覆盖塑料拉曼光谱（SI 附录，图 S1）的特征强特征，我们精心选择 793、804、886 和 897 nm 作为四个中心波长，以包括 C–H（不饱和碳和饱和碳，3,110 至 2,800 cm ⁻¹）、酯键（1,770 至 1,670 cm ⁻¹）和双键振动（1,660 至 1,580 cm ⁻¹）的强烈特征光谱特征，以便更好地区分每种塑料类型。我们通过测量七种最常见的塑料聚合物的体SRS光谱构建了一个小型库（图2A ）：聚酰胺66（PA）、聚丙烯（PP）、聚乙烯（PE）、聚甲基丙烯酸甲酯（PMMA）、聚氯乙烯（PVC）、聚苯乙烯（PS）和聚对苯二甲酸乙二醇酯（PET），具有精细的光谱间隔（~3cm ⁻¹）。

与体光谱测量不同，纳米塑料的单粒子成像需要更小的像素尺寸、更长的积分时间和更高的功率才能获得最佳信噪比。因此，由于检测灵敏度和特异性之间的根本权衡，几乎不可能以如此精细的光谱间隔（每个 FOV 的成像时间为数小时，在此期间样品漂移和燃烧的可能性不断增加）测量纳米塑料。此外，基于光谱聚焦的高光谱 SRS 显微镜的光谱分辨率通常为 10 至 25 cm ⁻¹。为了在吞吐量和光谱分辨率之间取得适当平衡的高效高光谱成像，我们进一步对光谱（SI 附录，图 S6）进行了子采样，光谱间隔为 ~15 cm ⁻¹，这仅略高于光谱分辨率并产生了可接受的成像吞吐量（每 0.2 mm × 0.2 mm FOV ~0.5 小时）用于纳米塑料的单粒子化学成像。

高通量塑料颗粒分析还需要自动光谱分析来识别塑料。基于 FTIR 或拉曼光谱的微塑料分析普遍采用用于自动化学识别的光谱匹配算法（40、41 ）。在典型的环境研究中，需要分析数千个颗粒光谱，手动塑料识别和计数不仅极其耗费人力，而且还容易受到人为偏见的影响（14、40-42 ）。自动颗粒分析有助于加快测量速度，分析更多颗粒，并确保普遍且公正的塑料识别。了解到环境科学对自动化的需求，我们首先在 FTIR 和拉曼分析中应用经典的库匹配算法，但发现它们与窄带 SRS 高光谱分析不太兼容。以从研磨 PA 标准制备的颗粒 A 中检测到的光谱为例（图2B ）。经过背景减法和数据归一化的光谱预处理后，颗粒 A 的光谱与聚酰胺的 SRS 特征清晰匹配。然而，当使用常见的光谱匹配算法 ( 42 )（例如皮尔逊相关系数 (PC) 或平方欧几里得余弦 (SEC) 测量）测量颗粒 A 与库中块体塑料标准的光谱相似性时，识别结果似乎难以捉摸（图 2D和E ）。在实际样品分析中，不应假设颗粒 A 应该属于库中的任何标准塑料，这意味着必须基于给定的阈值对每个塑料标准独立做出是或否判断。FTIR 或自发拉曼分析微塑料时常用的阈值是相似度测量值高于 0.7，这显然太低而无法识别颗粒 A。由于 PS 纳米颗粒可作为模型标准，我们首先尝试研究高光谱 SRS 成像下每种纳米塑料分析算法的相似性阈值。然后，可以根据至少 95% 的 PS 颗粒（PC 的相似性指数高于 0.75，SEC 的相似性指数高于 0.94）的四分位数来确定相似性阈值。然而，对于颗粒 A 而言，进行二元识别判断仍然具有挑战性，因为三种塑料聚合物（PA、PP 和 PVC）的相似性测量值在数量上非常接近，并且均高于阈值（图 2 D和E）。）。需要注意的是，不能简单地在所有标准中选出最佳得分，因为在实际样品分析中，A 完全有可能是非塑性材料。事实上，如果我们基于大肠杆菌所代表的生物质的模型标准光谱来模拟可能的非塑性 SRS 光谱，那么对于这两种算法，超过 95% 的光谱与 PA 标准在给定阈值上具有相似性测量（SI 附录，图 S12 a和b）。

我们认为，上述困难的主要原因在于检测灵敏度和特异性之间的权衡。强调化学特异性，自发拉曼光谱或其他宽带相干拉曼显微镜可以通过在大量拉曼振动模式之间分配光功率来覆盖扩展的光谱窗口（> 1,000 cm ^{−1 ）。丰富的光谱信息可以通过简单的算法实现化学识别}，但在有限的像素停留时间内，检测灵敏度会降低数千倍（43-45）。然而，在纳米塑料分析的背景下，检测颗粒信号是从振动光谱进行化学识别之前的前提。为了在实际吞吐量下测量尽可能小的塑料颗粒，最终只能在合理的 SNR 下检测到最强的拉曼特征。对于大多数塑料（本质上是有机聚合物）而言，最强的拉曼特征位于有限的 C–H 振动窗口内。在这种情况下，特定的化学物质识别需要算法在有限的光谱窗口内精确捕捉形状特征，这超出了传统光谱匹配算法的能力。此外，在对微小纳米颗粒进行成像时，信噪比不可避免地会受到影响和限制，这给光谱解释以进行可靠的化学识别带来了进一步的挑战。因此，需要新的方法来应对受激拉曼散射（SRS）仪器带来的特异性挑战，该仪器能够实现纳米塑料成像前所未有的灵敏度。

3. 数据驱动的SRS定制光谱匹配算法恢复化学特异性

利用数据科学，我们旨在开发算法来解释检测到的SRS特征的形状，并检索用于聚合物识别的化学特异性。首先，开发一个SRS定制光谱匹配系数（SMC _SRS）作为量化光谱相似性的指标，同时最小化噪声干扰（图2，公式1）。SMC _SRS使用一种优化算法，该算法考虑了检测到的SRS光谱源自缩放（强度因子α ）归一化本体标准光谱加上成像条件下的一定背景贡献（ β ，β< 1). 拟合光谱（ αβ ) 与检测到的粒子光谱进行了比较找到最小可能的光谱距离作为 SMC _SRS。较小的 SMC _SRS值表示与相应标准的光谱相似度较高。该指标 SMC _SRS为检测纳米塑料提供了多种优势。优化算法同时考虑所有光谱点，从而减少了噪声对每个特定光谱点的直接影响。拟合过程利用了相似性测量的可靠性。此外，测量结果是可解释的。明确定义的强度因子 α 和背景因子 β 可以指示来自每个光谱成分（颗粒和周围背景）的贡献。最后，光谱距离测量提供了度量相似性评估。

通过这种精细量化光谱相似性的方法，我们再次面临一个挑战：如何对聚合物进行非任意的二元判断。我们计划开发一种基于学习的方法，以确定此前难以确定的用于识别所有塑料聚合物的二元阈值。我们的前提是，如果我们能够测量谱库中所有类型塑料的纳米颗粒光谱，我们就能够从数据中学习，并根据已知身份的颗粒分布绘制出正确的识别边界。然而，实际上，市面上只有聚苯乙烯纳米球，其化学成分和纳米尺寸已得到充分表征。由于缺乏其他聚合物纳米颗粒的可靠基础数据，我们不得不寻找其他方法来收集精确阈值测定所需的海量信息。

受到人工智能中合成数据日益增多的实用性（46）以及数据科学在SRS显微镜中日益增长的参与度（47-49 ）的启发，我们意识到我们可以从块体标准光谱中模拟纳米塑料的实验SRS光谱，以用作训练数据集（即合成数据）。基于我们对SRS仪器的理解，我们提出了一个模型，其中典型的高光谱SRS光谱中有两个主要噪声源：一个是SRS强度上的基波噪声，就像散粒噪声限制场景中一样，可以从同一张SRS图像中轻松读出；另一个是SRS仪器造成的频率不确定性，其中激光轮廓和移动延迟台都可能导致每次测量中激发的实际频率在预设光谱点附近波动。假设波动遵循高斯分布，我们以PS纳米球作为标准模型来研究波动范围，并发现从PS纳米颗粒的合成光谱和测量光谱计算出的SMC _{SRS具有惊人的一致性（} SI附录、补充说明2和图S10 ）。噪声来源的组合性质解释了SMC _SRS值对光谱强度的 依赖性（α )，正如模拟所建议并通过实验验证（图 2 F）。

对库中的所有标准应用相同的模型，我们生成了一个合成数据集，其中包含塑料库中每种聚合物的纳米塑料可能的SRS光谱。SMC的_SRS值在粒子的光谱之间呈现出良好的分离。(  ，R 是标准聚合物的正确标识）和粒子光谱(  ) 在所有散点图中（SI 附录，图 S11）。利用大量生成的合成数据点，根据散点趋势拟合对数函数作为聚合物鉴定的阈值线（SI 附录，补充说明 2和表 S2）。

我们首先通过模拟库中所有标准的另一个合成数据集作为测试数据来评估识别性能。与传统的光谱匹配算法相比，针对 SRS 定制开发的算法在塑料识别中显示出最小的假阳性率（SI 附录，图 S12）。不超过 0.5% 的非塑料光谱（由大肠杆菌模拟）被误认为是库中任何塑料类型的光谱（SI 附录，图 S12 c），这比使用传统光谱匹配算法超过 97% 的光谱（SI 附录，图 S12 a和b）有了显著的改善。相似 SRS 光谱的聚合物之间的假阳性也大大减少，最大值约为 5% PA 被误认为 PP（SI 附录，图 S12 c）。如果使用 PC 或 SEC 作为具有确定阈值的相似性测量，则相同的数字也高达 97% 以上（SI 附录，图 S12 a和b）。

为了进一步解决可能出现的罕见情况，即一个粒子被鉴定为库中多种聚合物的匹配物，相应粒子的化学身份将分配给具有最小 SMC _SRS值的聚合物。通过已建立的光谱识别工作流程，对于库中的所有聚合物，都可以实现超过 96% 的识别率，假阳性率低于 1%（SI 附录，图 S12 d）。由于 PS 纳米球是唯一可用的纳米塑料标准，因此该工作流程的实验验证基于用低温磨机研磨聚合物标准品制备的相应微塑料的成像。为了最大程度地模拟相似水平的光谱变化，相应地调整成像条件以匹配纳米塑料测量的信噪比。最后，我们证实了实验粒子测量中的相同识别率超过 96%，并且没有观察到将塑料粒子误认为是库中的其他聚合物（图 2 G）。

这种数据驱动算法的开发允许在受限的光谱窗口内识别具有不同振动特征的每种塑料聚合物，从而检索自动光谱识别所需的化学特异性。重新审视颗粒 A 和标准 PS 纳米球 B 的识别，我们可以在整个库中正确识别颗粒 A 和颗粒 B 为 PA 和 PS（图 2 H- N），SMC _SRS很好地捕获了传统算法遗漏的形状差异和从数据驱动研究中学习到的阈值。将数据科学的思维方式与先进的测量科学相结合，我们最终克服了高通量高光谱 SRS 分析的基本灵敏度-特异性权衡。同时实现了窄带 SRS 扩增的卓越纳米灵敏度和具有强大化学识别的化学特异性，以填补纳米塑料分析工具中缺失的空白。

4. 开发瓶装水中微纳米塑料检测工作流程

平台建立后，我们开始应用该实用程序研究实际样品中的微纳米塑料。微塑料广泛存在于人类食品 ( 50 )、饮料 ( 51 ) 和产品包装 ( 52 – 55 ) 中，其中瓶装水尤其令人关注，因为它是日常生活中摄入微塑料的重要来源 ( 56 – 59 )。受分析科学中灵敏度-特异性权衡的限制 ( SI 附录，图 S18 b )，文献知识仅限于瓶装水中的微塑料 ( SI 附录，表 S4 ) ( 19 , 60 – 62 ) ，而纳米塑料大多尚未被探索。到目前为止，仅报道了使用多种技术组合进行集成表征以分析瓶装水中的浓缩纳米颗粒等分试样。需要获取信息以解决单颗粒水平上纳米塑料污染的内在异质性（SI 附录，图 S18 a）（63、64 ）。本文报告了一种简洁的工作流程，该工作流程通过利用受激拉曼散射 (SRS) 显微镜进行快速单颗粒化学成像并实现纳米灵敏度，从而实现对微纳米塑料的全面表征。单次测量即可获取丰富的信息，从而同时表征化学成分和形貌，并通过高通量单颗粒分析实现多维统计分析。

过滤是将超过一定尺寸的颗粒收集到膜表面的最常用方法之一。如果收集的膜可直接用于 SRS 成像，则对于分析实际样品将非常有利。氧化铝膜在目标光谱窗口中具有最小的背景，并表现出与振动光谱良好的兼容性。通过施加重水以减少折射率不匹配，可以轻松地将看似不透明的氧化铝膜转变为透明的成像窗口。这导致了具有可接受信号保留的透射 SRS 成像（~70% 的原始灵敏度，SI 附录，图 S7 b和c _）。用D2O制备的琼脂糖凝胶将颗粒原位嵌入膜表面，进一步实现了单个颗粒的稳态 SRS 成像，并将成像背景降至最低。通过这种方式，简洁的样品预处理足以对原始过滤膜进行高质量的 SRS 成像（SI 附录，图 S7 a），避免在任何复杂的样品干燥或转移过程中出现不良的样品损失或污染。

图 3显示了使用高光谱 SRS 成像分析瓶装水中微纳米塑料暴露的既定工作流程。对于每个样品，在收集区域内随机采样五个或更多视野 (FOV)，以便在 SRS 显微镜下进行高光谱成像（图 3D ）。在每个 FOV 中，通过集成数据分析工作流程检测微纳米塑料，该工作流程使用开发的算法和验证的阈值条件自动执行粒子分割和塑料识别。然后结合从高光谱 SRS 图像获得的每个塑料颗粒的形态和化学信息，提供高维分析（图 3E ）。按照该程序，我们分析了同时从一家大型零售商处采购的三个不同品牌的瓶装水。由于实验室中无法获得无塑料水（SI 附录，补充说明 6），Anodisc 过滤器的制备和测量方式与空白对照相同。结果显示，我们能够通过与相应的本体标准的光谱匹配，明确地检测出库中所有七种塑料聚合物的单个颗粒（图4），这证明了我们数据驱动的高光谱SRS成像平台强大的塑料识别能力。

5. 瓶装水中微纳米塑料的多维分析

通过对已识别的塑料聚合物成分的单粒子图像进行量化，可以提供多维信息，以构建瓶装水中未充分探索的纳米塑料的分析全景。

通过颗粒计数进行数量量化表明，三个不同品牌的每个视场 (FOV)（0.2 毫米 × 0.2 毫米）内平均可识别出 78 至 103 个塑料颗粒，显著高于（P < 0.001）空白样品（图 5A ）。假设微纳米塑料颗粒在膜区域表面均匀分布（SI 附录，补充说明 5），我们可以估算瓶装水中的微纳米塑料暴露量。我们估计平均约有 2.4 ± 1.3每升不同品牌瓶装水中测量到的塑料颗粒摄入量约为105个^（图5C ）。对每种聚合物的单个颗粒进行单独分析，以揭示其化学异质性。在数据库中，PA、PP、PET、PVC和PS可能在瓶装水中的微纳米塑料暴露中发挥重要作用（图5B ）。不同品牌的微纳米塑料的具体化学成分各不相同，但在我们分析的三个品牌中，PA似乎是数量上共同的主要贡献者。

利用SRS强度和焦点体积内分析物数量之间的线性关系，我们除了能够估算粒子数量之外，还能估算质量暴露量。可以通过标准PS纳米球获得的线性关系（SI附录，图S16），根据密度和相对SRS强度估算每种聚合物的质量校准曲线。这样，每个粒子在感兴趣区域内的积分强度就转化为质量（图5E和F ）。估计的微纳米塑料质量暴露量约为10ng/L。分析质量中的化学成分，我们发现质量贡献和数量贡献之间存在不可忽视的差异。以品牌C的结果为例，PS纳米塑料虽然在粒子数量上占主导地位，但只占质量的一小部分。相反，PET成为质量的主要贡献者。这种表面上的差异凸显了从集体粒子表征角度对塑料成分的潜在误解，这种误解源于现实世界样本中微纳米塑料的异质性。

利用 SRS 显微镜对单个颗粒进行形态表征可直接揭示颗粒异质性的另一个维度。本文报告了对具有明确身份的单个微纳米颗粒图像的粒度和形状进行统计分析的结果。测量粒度分布时，我们能够通过从强度读数外推粒度（假设颗粒为实心球体）并使用颗粒体积与 SRS 信号之间的线性关系作为校准（SI 附录，补充说明 3）来表征衍射极限以下的颗粒。结果我们发现不同化学成分的塑料颗粒实际上具有不同的粒度分布模式（图 6 A– G）。这里对颗粒异质性的直接观察为从质量或数量测量中观察到的化学成分差异提供了自然的解释。以PS和PET为例，PS颗粒的尺寸分布集中在100~200nm，而PET颗粒的尺寸分布则趋向于1~2微米，这也就解释了为什么在质量测量时PET是更重要的成分，而在计算颗粒数量时PS明显占主导地位（图5D和F ）。

形状是另一个重要的形态特征，是纳米毒性的一个关键方面。研究表明，形状在决定微纳米颗粒的细胞摄取方面起着重要作用（65，66 ）。塑料颗粒的SRS图像证实了瓶装水中微纳米塑料的形状多样性。为了以统计的方式解释塑料颗粒的形状，我们测量了衍射极限以上的单个颗粒的长宽比（图6H ）。长宽比在纳米毒理学研究中得到广泛认可（67，68）。检测到的塑料颗粒的长宽比范围从1到6，颗粒的平均长宽比约为1.7。图6I - M以图形方式显示了长宽比与颗粒形状之间的关系。长宽比大于3的颗粒最有可能是纤维状的，而长宽比低于1.4的颗粒则基本上是球形的。所有检测到的聚合物中都发现了塑料颗粒形状的变化，这证实了人们普遍认可的观点，即现实世界中的微纳米塑料具有多种形态特性。这种尺寸很难与实验室中常见的工程聚合物纳米颗粒相媲美，而且现实生活中接触塑料颗粒及其不同的物理化学特性（即尺寸、形状）所带来的毒理学后果尚待确定。

6.讨论和结论

通过开发用于微纳米塑料分析的数据驱动高光谱 SRS 成像平台，我们描述了一种提高纳米颗粒检测灵敏度和聚合物识别特异性的方法，这使我们能够开始解决纳米塑料的长期知识空白。我们估计，普通瓶装水中的微纳米塑料暴露量为每升 105 个颗粒，比以前报告的仅关注大型微塑料的结果（SI 附录，表 S4）（58、59、61、69、70）高出两到三个数量级^。就人类暴露量估计而言，这些值大大高于目前文献中报告的值（56、71 ） ，这是由于新检测到的塑料颗粒中的纳米塑料部分造成的。以前在传统成像下看不见的微小颗粒实际上在数量上占主导地位，约占检测到的全部塑料颗粒总数的 90%。其余 10% 被鉴定为微塑料的颗粒浓度约为每升3 × ^{104 个颗粒（} SI 附录，图 S17），其中大多数颗粒的尺寸在 2 µm 以下。较大的颗粒（> 2 µm）在常规光学显微镜下更容易识别，根据基于不同技术报告的检测限，它们与报告的微塑料分析处于同一数量级（SI 附录，图 S17 和表 S4）。我们的结果通过明确检测实际样品中的纳米塑料证实了微米级以上的塑料碎片。与自然界中的许多其他粒度分布类似，尽管在传统的粒子成像技术下看不见或无法识别，但纳米塑料的数量远远多于以前计算出的大型微米塑料。由于尺寸较小的纳米颗粒含有非立方体物质，因此在质量定量中这种纳米塑料群体也很容易被忽视。然而，考虑到这些纳米塑料颗粒穿过生物屏障的能力，尽管纳米颗粒对质量测量的贡献看似微不足道，但可能在毒性评估方面发挥主导作用（72,73）。

我们还发现许多检测到的颗粒的 SRS 光谱与任何标准都不匹配。事实上，我们由七种塑料聚合物组成的小型库只能占 SRS 显微镜下成像的总颗粒/点的约 10%。使用振动显微镜对瓶装水中的微塑料进行分析时也报告了类似的识别率，表明看似简单的水样内部的颗粒组成复杂（SI 附录，表 S4）。从这个意义上讲，如果我们假设所有检测到的有机颗粒都来自塑料[SEM-EDX 或尼罗红染色的定量结果也包含相同的假设（19 , 74）]，则微纳米塑料浓度可能高达每升 106 个^颗粒。然而，天然有机物的普遍存在当然需要与具有聚合物特异性的光谱进行谨慎区分。此外，对未识别颗粒的仔细研究表明其他方面进一步增加了识别化学成分的复杂性。例如，一些颗粒在指纹区域中表现出与 PET 特征两个峰（C=O 酯键：1,730 cm ⁻¹；C=C 双键：1,615 cm ⁻¹）相同的特征，但在高频 C–H 区域呈现出各种各样的振动峰（SI 附录，图 S8 a– d）。不同于 PET 的聚合物材料不太可能显示与标准 PET 光谱完全匹配的 C=O 和 C=C 振动特征。更合理的解释是，它们是包含 PET 和其他成分的小异质聚集体，其 SRS 光谱是来自每个成分光谱的叠加。事实上，对于一些较大的聚集体，我们甚至可以捕捉到聚集体内的空间化学异质性（SI 附录，图 S8 a、e和i）。长期以来，纳米塑料或其他天然有机物之间可能形成异质聚集体，这被认为是纳米塑料分析中的一个潜在挑战，并且可能影响生物暴露中的毒理学结果( 11 )。在实际样品中直接可视化此类异质聚集体证实了这种担忧。对于其他可能在未使用PET的情况下形成的异质聚集体，严格的识别需要扩展光谱库，并改进SRS显微镜或其他具有扩展光谱窗口的振动成像技术的分析算法，以应对大量颗粒异质性带来的挑战( 27,75,76 ) 。

另一个重要的发现是，不同的化学组成会导致粒度分布发生变化，这表明颗粒形貌与化学组成之间存在相互联系。观察到的塑料组成与颗粒形貌之间的非正交性挑战了通过集合测量表征微纳米塑料的传统假设。以品牌C的分析结果为例，微纳米塑料的集合测量可能表明，从组成分析来看主要物质是PET，而从形貌分析来看，大多数塑料颗粒的尺寸小于500纳米。假设这两个维度是独立的属性，人们可能会认为品牌C瓶装水中的大多数塑料颗粒应该是尺寸小于500纳米的PET颗粒。然而，我们的单颗粒分析结果却呈现出明显的差异：样品中含有少量约为微米尺寸的PET颗粒和大量尺寸小于500纳米的PS颗粒。

这种非正交性可能提供有价值的信息，以理解、追踪并最终预防微纳米塑料污染的潜在来源。具体来说，在饮用水生产过程中，从水井到水瓶的每个步骤中都存在塑料污染（77）。在不同塑料聚合物之间发现的尺寸差异可能为水生产过程中的污染源提供宝贵的信息。例如，我们分析的三个品牌瓶装水的包装材料PET和PE具有相似的尺寸分布模式，与其他聚合物相比，微米级颗粒占主导地位。一种可能的解释是，一些此类颗粒是在运输或储存过程中从瓶装包装中新释放出来的，它们被完整地保留在水样中。其他聚合物，如PA、PP、PS和PVC，虽然不是包装材料，但也被识别出数量可观，很可能是在水生产之前或过程中引入的。PP和PA具有相同的广泛尺寸分布，被广泛用作水处理中的设备组件或助凝剂（78）。具体来说，PA 是反渗透中最常用的膜材料 ( 79 )，而反渗透是这三个品牌共用的常见水净化方法。PVC 和 PS 具有独特的尺寸分布，有利于小纳米塑料，可能更早地指示污染源。微塑料分析确定 PVC 是原水中最丰富的聚合物类型 ( 77 )。众所周知，PS 是水净化中离子交换树脂的骨架材料 ( 80 )。在水处理的后期步骤中，大颗粒的 PVC 或 PS 可能会被 RO 膜去除，留下的大部分是纳米颗粒。

最后，颗粒形貌与化学成分之间的相互关联对毒理学问题具有深远的影响。正如工程纳米颗粒的研究表明以及塑料颗粒研究开始表明的那样，微纳米颗粒引起的毒性不仅与剂量有关，还与颗粒的物理化学特性及其对细胞相互作用和吸收的影响有关（81,82 ） 。以C品牌瓶装水为例，PS纳米塑料加上少量PET微塑料引起的细胞毒性可能与PET纳米颗粒的假定效应不同。真正全面的微纳米塑料毒性评估需要对塑料颗粒进行多维表征，并整合每个塑料颗粒在化学成分和颗粒形貌上的不同特性。具有纳米颗粒灵敏度和塑料特异性的单颗粒成像为解决日益严重的毒性问题提供了不可或缺的信息。它不仅能够对塑料颗粒进行精确的暴露量化分析，而且还具有直接可视化颗粒与生物相互作用的独特潜力。因此，我们设想，数据驱动的高光谱SRS成像平台将继续弥合纳米级塑料污染的知识空白，并利用扩展的光谱库来研究更复杂的生物和环境样本。

7.材料和方法

致谢

我们感谢数据科学家 Tingran Wang 和 Mariam Avagyan 对算法的讨论。我们感谢哥伦比亚大学科学与工程研究计划 (RISE)、哈德逊河基金会、美国国立卫生研究院 (NIEHS) 北曼哈顿环境健康与正义中心 (NIEHS P-30-ES009089) 以及罗格斯大学环境暴露与疾病中心 (NIEHS P30-ES005022) 的支持。

1. SRS Imaging of Polystyrene Nanospheres with Single-Particle Sensitivity

SRS microscopy is well known to be orders of magnitude faster than regular Raman imaging (25, 26). The drastically higher imaging speed of SRS microscopy hence provides high throughput on particle imaging. However, whether high-speed SRS has a better detection limit than regular Raman and whether it can actually reach the single-particle sensitivity of nanoplastics are not obvious. A theoretical quantification is helpful to address the question in the first place. For a given major type of plastic polymer, we can estimate the mass of a 100-nm-diameter nanoplastics based on the plastic density and calculate the number of repeating units (i.e., constituting monomer) via its molecular weight. As shown in SI Appendix, Table S1, this number is around 10⁶ for most major plastic types, based on which we can further estimated the number of most abundant chemical bonds in a single plastic particle to be ~10⁷.

We can then theoretically explain why a 100 nm nanoplastic particle is difficult to be detected by conventional Raman microscopy. The spontaneous Raman cross-section of a typical C–H vibration is about 10⁻²⁹ cm². Hence, the spontaneous Raman cross-section of a 100-nm nanoparticle is 10⁻²² cm². The laser waist area can be shrunk to about 2 × 10⁻⁹ cm² under a high numerical aperture microscope objective. The probability of Raman scattering event per excitation photon is then (10⁻²² cm²)/(2 × 10⁻⁹ cm²) = 5 × 10⁻¹⁴. Assuming a moderately high laser power of 10 mW with a conventional 532 nm laser, which corresponds to an excitation flux of 3 × 10¹⁶ photons/s, and a rather long acquisition time of 100 ms (a small 128 × 128 image will take half an hour), only about 130 photons can be generated per particle in total via spontaneous Raman scattering. Considering the quantum yield of the entire instrument (including objective, filters, pinhole, spectrometer, and camera) typically is ~1%, roughly only 1.3 photons can be ultimately detected. Such a feeble signal can be easily overwhelmed by noise from other backgrounds such as autofluorescence.

By employing an additional coherent Stokes laser, SRS amplifies the feeble scattering crossing section of a specific spectral mode (defined by the energy difference between pump and Stokes lasers) via quantum stimulation. When a pulsed narrowband Stokes laser is used (24, 33), the stimulated Raman enhancement factor can be maximized to more than 10⁸ (32, 34). The probability of a stimulated Raman scattering event per pump excitation photon then becomes 5 × 10⁻⁶, which is measured as a stimulated Raman loss experienced by the pump beam targeting C–H vibration. The noise of the pump beam under high-speed SRS microscopy acquisition (18 µs/pixel) is measured to be 5 × 10⁻⁷ (Fig. 1), which is about 10× lower than the expected stimulated Raman loss signal from a single 100-nm plastic particle. Thus, we predict that narrowband SRS shall break the detectability barrier of spontaneous Raman and bring a single nanoplastic particle into detection in just tens of microseconds.

We then experimentally verify the detection sensitivity using standard plastic particles. Polystyrene is one of the most common plastics widely used in daily life. Polystyrene particles of specified sizes are commercially available as analytical standards and have been routinely used as a model material to study micro-nanoplastics (35, 36). The Raman spectrum of polystyrene suggests a prominent peak at 3,050 cm⁻¹ from aromatic C–H vibration on the phenyl ring (SI Appendix, Fig. S1), which can be selectively amplified for SRS imaging by tuning the difference of pump and Stokes beams to match this transition energy. Using commercial PS micro-nano spheres from 100 nm to 3 µm, we evaluated the detection sensitivity of our SRS microscope in imaging nanoplastics. To stabilize the particles during imaging, we embedded the diluted PS particles in agarose gel. As the particle size goes smaller, the residue of the water background around 3,000 cm⁻¹ starts to dominate (SI Appendix, Fig. S2a), overwhelming the authentic spectrum of individual PS nanoparticles. To resolve this background issue for better imaging contrast, we substituted regular H₂O with D₂O to prepare the agarose gel (SI Appendix, Fig. S2b). Compared to H₂O, the Raman spectrum of D₂O is red-shifted to the silent region (2,200 to 2,800 cm⁻¹, SI Appendix, Fig. S3), creating a background-free environment for probing C–H vibration.

SRS intensity of individual particles can be thereby measured from single-channel narrow-band imaging with high-throughput (~1,000 particles in one 51 × 51 µm FOV within 2 s, SI Appendix, Fig. S4). This imaging speed is orders of magnitude faster than other nanoplastic imaging techniques, such as AFM-IR and STXM (21, 23, 37). With the optical diffraction limit, the optimal spatial resolution of SRS microscopy is measured to be 365 nm (Fig. 1 H and I). With a spatial sampling of 200 nm pixel size for high-throughput imaging, individual PS nanospheres of above 500 nm can be discerned with their shape from the images (Fig. 1 D–G). When the size of the particles goes smaller than the diffraction limit (Fig. 1 A–C), the finite optical resolution renders the particle image a diffraction-limited pattern. Yet, the SRS intensity of a single particle can still be readily recognized down to 100 nm based on the diffraction limit pattern and the intensity distribution (SI Appendix, Fig. S5). Thus experimentally, we have shown that compared to regular spontaneous Raman, SRS imaging can offer orders of magnitude higher imaging speed/throughput and a superior limit of detection for nanoplastics analysis.

A linear relationship was observed between the logarithm of SRS signal (  ) and the logarithm of diameter for PS particles smaller than 0.7 µm (Fig. 1J and SI Appendix, Supplementary Note3). The trendline with a slope of 2.98 within the range indicates the SRS signal (  ) increase linearly with the particles’ volume, which scales in cubic as the particles’ diameters increase. When the particles’ size is enlarged to overfill the effective focal volume sequentially in first x, y, and later z dimensions (SI Appendix, Fig. S14), the linear dependency disappears. This good linearity (R² = 0.998) is due to the fundamental linear dependency of the SRS signal on the concentration of the target analyte, providing powerful utilities in several aspects. First, the actual size of particles below the diffraction limit can be estimated based on the obtained calibration curve (SI Appendix, Fig. S16a), extending the size characterization limit. Second, with the known information on the plastic density, the same calibration curve can be transformed into a reference to deduce a particle mass out of a detected SRS nanoplastics image (SI Appendix, Supplementary Note3 and Fig. S16b). Finally, taking an SNR of one as the threshold, the detection limit of our narrowband SRS microscope can be determined (Fig. 1K) to reach PS nanospheres down to 60 nm.

2. Fundamental Challenges on Chemical Identification of Nanoplastics with Hyperspectral SRS Imaging

Nano-sensitivity solves the first-order issue to ensure the plastic particles are detectable. The chemical specificity of a technique is also crucial to identify plastics from other co-existing substances and further distinguishing plastic polymers from each other. Harnessing vibrational spectroscopy as imaging contrast, SRS microscopy, in principle, holds the demanded specificity for chemical imaging. Instrumentally, we perform hyperspectral SRS imaging via the spectral-focusing technique (38, 39). To best cover the characteristic strong feature of the plastic Raman spectrum (SI Appendix, Fig. S1) within the tuning range of the instrument (790 to 910 nm), we carefully choose 793, 804, 886, and 897 nm as four central wavelengths to include the strong and characteristic spectral features of C–H (unsaturated and saturated carbons, 3,110 to 2,800 cm⁻¹), ester bonds (1,770 to 1,670 cm⁻¹), and double bond vibration (1,660 to 1,580 cm⁻¹) for better distinguishment between each plastic type. We constructed a small library by measuring the bulk SRS spectra of seven most common plastic polymers (Fig. 2A): polyamide 66 (PA), polypropylene (PP), polyethylene (PE), polymethyl methacrylate (PMMA), polyvinyl chloride (PVC), polystyrene (PS), and polyethylene terephthalate (PET) with fine spectral intervals (~3 cm⁻¹).

Unlike bulk spectra measurement, single-particle imaging of nanoplastics requires a much smaller pixel size, longer integration time, and higher power for optimal signal-to-noise ratio. Therefore, due to the fundamental trade-off between detection sensitivity and specificity, it is nearly impossible to measure nanoplastics with such fine spectral intervals (hours of imaging time per FOV with increasing possibility of sample drifting and burning during the time). Moreover, the spectral resolution of a hyperspectral SRS microscope based on spectral focusing is typically 10 to 25 cm⁻¹. For efficient hyperspectral imaging with a proper balance between throughput and spectral resolution, we further subsampled the spectra (SI Appendix, Fig. S6) with the spectral interval of ~15 cm⁻¹, which is only slightly above the spectral resolution and yielded acceptable imaging throughput (~0.5 h per 0.2 mm × 0.2 mm FOV) for single-particle chemical imaging of nanoplastics.

High-throughput plastic particle analysis also requires automated spectral analysis for plastic identification. Spectral matching algorithms for automated chemical identification are prevalently adopted in microplastic analysis based on FTIR or Raman spectroscopy (40, 41). With thousands of particle spectra in need of analysis in a typical environmental study, manual plastic identification and counting are not only impossibly labor-intensive but also subjected to human bias (14, 40–42). Automated particle analysis helps to speed up the measurement, analyze more particles, as well as ensure ubiquitous and unbiased plastic identification. Understanding the need for automation in environmental science, we started with applying the classic library matching algorithms in FTIR and Raman analysis but found them not so compatible with narrow-band SRS hyperspectral analysis. Take a detected spectrum from particle A prepared from grinding the PA standard as an example (Fig. 2B). After spectrum pre-processing on background subtraction and data normalization, the spectrum of particle A clearly matches the SRS signature of polyamide. However, when measuring the spectral similarities of particle A to bulk plastic standards from the library using common spectral matching algorithms (42), such as Pearson’s correlation coefficient (PC) or squared Euclidean cosine (SEC) measurement, the identification results appears elusive (Fig. 2 D and E). In a real-life sample analysis, there should be no premise to assume particle A should belong to any standard plastics in the library, which means a yes or no judgment has to be made independently for each plastic standard based on a given threshold. The common threshold employed in FTIR or spontaneous Raman analysis of microplastics is the similarity measurement above 0.7, which is clearly too low to identify Particle A. Since PS nanoparticles are available as model standards, we first try to study the similarity threshold of each algorithm for nanoplastics analysis under hyperspectral SRS imaging. The similarity threshold can then be determined based on the quartile of identifying at least 95% of the PS particles (similarity index above 0.75 for PC, and similarity index above 0.94 for SEC). However, the challenging part of making a binary identification judgment remains in the case of particle A as similarity measurements from three plastic polymers (PA, PP, and PVC) are very close in number and all above the threshold (Fig. 2 D and E). Note that one cannot simply pick the best score among all the standards because it is totally possible for A to be nonplastic materials in real sample analysis. In fact, if we simulate the possible nonplastic SRS spectra based on the model standard spectrum of biomass represented by E. coli, over 95% of them will have similarity measurements against PA standard over the given threshold for both two algorithms (SI Appendix, Fig. S12 a and b).

We reflect that the main reason underlying the above difficulty stems from the trade-off between detection sensitivity and specificity. Emphasizing the chemical specificity, spontaneous Raman spectroscopy, or other broadband coherent Raman microscopy can cover an extended spectral window (>1,000 cm⁻¹) by distributing the optical power among a large number of Raman vibrational modes. The rich spectral information can enable chemical identification with simple algorithms but comes with the cost of over thousand times compromised detection sensitivity under a limited pixel dwell time (43–45). However, in the context of nanoplastics analysis, detecting the particle signal is the premise before chemical identification from the vibrational spectrum. With the aim of measuring as small plastic particles as possible under practical throughput, eventually, only the strongest Raman features will be detectable with reasonable SNR. For most plastics, which are organic polymers by nature, the strongest Raman signatures reside within the limited C–H vibration window. In this case, specific chemical identification requires the algorithms to precisely capture the shape feature within the restricted spectral window, which is beyond the capacity of conventional spectral matching algorithms. Moreover, the inevitably compromised and circumscribed signal-to-noise ratio when imaging diminutive nanoparticles create further challenges in spectral interpretation for robust chemical identification. Therefore, new methods are demanded to address the specificity challenge imposed by the SRS instrumentation that enables unprecedented sensitivity in imaging nanoplastics.

3. Data-Driven SRS-Tailored Spectral Matching Algorithm Recovers Chemical Specificity

Harnessing data science, we aim to develop algorithms to interpret the shape of detected SRS features and retrieve the chemical specificity for polymer identification. First, an SRS-tailored spectral matching coefficient (SMC_SRS) is developed as an indicator to quantify spectral similarity with minimized noise interference (Fig. 2, eq. 1). SMC_SRS uses an optimization algorithm that considers the detected SRS spectrum  originating from scaling (intensity factor α ) the normalized bulk standard spectrum  plus a certain background contribution at the imaging condition ( β , β < 1). The fitted spectrum ( αβ ) was compared with the detected particle spectrum  to find the minimum possible spectral distance as SMC_SRS. The smaller SMC_SRS value indicates a higher spectral similarity to the corresponding standards. This indicator SMC_SRS provides several advantages for the purpose of detecting nanoplastics. The optimization algorithm considers all spectral points simultaneously, which reduces the direct influences induced by the noise on each particular spectral point. The fitting process leverages the reliability of the similarity measurement. In addition, the outcome of the measurement is interpretable. The well-defined intensity factor α and background factor β can indicate the contribution from each spectral component (the particle and the surrounding backgrounds). Finally, the spectral distance measurement provides metric similarity evaluation.

With the spectral similarity quantified in this refined way, we returned to face the challenge of making a nonarbitrary binary judgment for polymer identification. We planned to develop a learning-based method to determine the previously elusive binary threshold for the identification of all plastic polymers. Our premise is that if we can measure the nanoparticle spectra for all types of plastics within the library, we shall be able to learn from the data and draw the correct boundary for identification based on the distribution of the particles with known identities. However, in reality, only PS nanospheres are commercially available to us with well-characterized chemical composition and nano sizes. Without reliable ground truth from other polymer nanoparticles, we have to seek alternative ways to gather the massive information needed for rigorous threshold determination.

Inspired by the increasing utilities of synthetic data in AI (46), and the growing involvement of data science in SRS microscopy (47–49), we realized that we could simulate the experimental SRS spectra of nanoplastics from the bulk standard spectra to serve as a training dataset (i.e., synthetic data). Based on our understanding of the SRS instrumentation, we proposed a model, where there are two main sources of noise in a typical hyperspectral SRS spectrum: one is fundamental noise on the SRS intensity as in a shot-noise-limited scenario, which can be easily read out from the same SRS image; the other is the frequency uncertainty imposed by the SRS instrumentation, where both the laser profile and the moving delay stage can result in fluctuation of the actual frequency excited in each measurement around the preset spectral points. Assuming the fluctuation follows a Gaussian distribution, we used PS nanospheres as the standard model to investigate the fluctuation range and found an impressive consistency in SMC_SRS calculation from the synthetic spectra and measured spectra of PS nanoparticles (SI Appendix, Supplementary Note 2 and Fig. S10). The combinatory nature of noise origins explains the dependency of the SMC_SRS value on the intensity of the spectrum ( α ), as suggested in the simulation and validated by the experiment (Fig. 2F).

Applying the same model for all standards in the library, we generated a synthetic dataset containing the possible SRS spectra for nanoplastics of each polymer in the plastic library. A nice separation of the SMC_SRS value appears between the spectra of particle  (  , R is the correct identity of standard polymer) and spectra of particle  (  ) in all scatter plots (SI Appendix, Fig. S11). With the massively generated synthetic data points, a logarithmic function was fitted according to the trend of the scattered points as the threshold line for polymer identification (SI Appendix, Supplementary Note 2 and Table S2).

We first evaluate the identification performance by simulating another synthetic dataset from all standards in the library as testing data. Compared with conventional spectral matching algorithms, the SRS-tailored developed shows minimal false positives in plastic identification (SI Appendix, Fig. S12). No more than 0.5% of nonplastic spectra (simulated from E. coli) is misidentified as a hit for any plastic types in the library (SI Appendix, Fig. S12c), which is a drastic improvement from over 97% using conventional spectral matching algorithms (SI Appendix, Fig. S12 a and b). False positive between polymers of similar SRS spectrum is also much reduced with the maximum to be around 5% PA misidentified as PP (SI Appendix, Fig. S12c). The same number is also as high as over 97% if PC or SEC are used as similarity measurements with the determined thresholds (SI Appendix, Fig. S12 a and b).

To further address the possible rare cases where a particle is identified as hits for more than one polymer in the library, the chemical identity of the corresponding particle will be assigned to the polymer with the smallest SMC_SRS value. With the established spectral identification workflow, an over 96% identification rate can be achieved with a false positive rate below 1% for all polymers in the library (SI Appendix, Fig. S12d). Since PS nanosphere was the only available nanoplastic standard, the experimental validation of the workflow is based on the imaging of the corresponding microplastics prepared from grinding the polymer standards with the cryo-mill. Hoping to mimic a similar level of spectral variation to the best extent, the imaging condition is adjusted accordingly to match the signal-to-noise ratio of nanoplastic measurement. Finally, we confirmed the same identification rate of over 96% in the experimental particle measurement with no observed plastic particles misidentified as other polymers within the library (Fig. 2G).

Development of this data-driven algorithm allows for the identification of each plastic polymer with distinct vibrational features in a restricted spectral window, thus retrieving the required chemical specificity for automated spectral identification. Revisiting the identification of particle A and standard PS nanosphere B, we can correctly identify both particle A and particle B across the library to be PA and PS (Fig. 2 H–N), with SMC_SRS well captures the shape differences missed by conventional algorithms and threshold learned from the data-driven study. Coupling the mindset from data science with advanced measurement science, we finally overcome the fundamental sensitivity-specificity trade-off for high throughput hyperspectral SRS analysis. Superb nano-sensitivity from narrow-band SRS amplification and chemical specificity with robust chemical identification are simultaneously accomplished to fill the missing void in tools for nanoplastics analysis.

4. Developing Workflow for Micro-Nano Plastic Detection from Bottled Water

With the platform established, we moved on to apply the utility to study micro-nano plastics from real-life samples. Microplastics have been widely found in human foods (50), drinks (51), and product packaging (52–55), among which bottled water is of particular interest for being an important source of microplastics to be ingested in daily life (56–59). Limited by the sensitivity-specificity trade-off in analytical science (SI Appendix, Fig. S18b), the literature knowledge is constrained to microplastics in bottled water (SI Appendix, Table S4) (19, 60–62), leaving the nanoplastics mostly uncharted. So far, only ensemble characterizations using combinations of techniques are reported to analyze the aliquots of concentrated nanoparticles from bottled water. Information is demanded to address the intrinsic heterogeneity of nanoplastics contamination at a single-particle level (SI Appendix, Fig. S18a) (63, 64). Here, we report a concise workflow for comprehensive micro-nano plastics characterization enabled by rapid single-particle chemical imaging with nano-sensitivity by SRS microscopy. Rich information can be acquired from a single measurement to achieve simultaneous characterization of chemical composition and morphology, enabling multi-dimensional statistics through high-throughput single-particle analysis.

Filtration is one of the most common methods to collect particles above certain sizes onto a membrane surface. It would be highly preferable for analyzing real-world samples if the collected membrane is directly compatible for SRS imaging. Aluminum oxide membranes have minimal background in the target spectral window and have shown good compatibility with vibrational spectroscopy. The seemingly opaque aluminum oxide membrane can be easily transformed into a transparent imaging window by applying heavy water to reduce refractive index mismatch. This resulted in transmissive SRS imaging with acceptable signal retention (~70% of the original sensitivity, SI Appendix, Fig. S7 b and c). Embedding the particles on the membrane surface in situ with agarose gel prepared with D₂O further enabled stationary SRS imaging of individual particles with minimal imaging background. In this way, a concise sample preprocessing is enough for high-quality SRS imaging of the original filtration membrane (SI Appendix, Fig. S7a), avoiding undesirable sample loss or contamination in any complicated sample drying or transferring processes.

The established workflow for analyzing micro-nano plastics exposure from bottled water with hyperspectral SRS imaging is presented in Fig. 3. For each sample, five or more fields of views (FOVs) were randomly sampled within the collecting area for hyperspectral imaging under SRS microscopy (Fig. 3D). In each FOV, micro-nano plastics were detected by an integrated data analysis workflow that automatically performed the particle segmentation and plastic identification with the developed algorithms and validated threshold conditions. Morphological and chemical information of each individual plastic particle obtained from the hyperspectral SRS images was then combined to provide high-dimensional profiling (Fig. 3E). Following the procedure, we analyzed bottled water from three different brands acquired at the same time from a large retailer. With no access to plastic-free water in the lab (SI Appendix, Supplementary Note 6), the Anodisc filters are prepared and measured in the same way as blank control. In the results, we were able to detect individual particles for all seven plastic polymers in the library unambiguously by spectral matching with their corresponding bulk standards (Fig. 4), demonstrating the powerful plastic identification capability of our data-driven hyperspectral SRS imaging platform.

5. Multidimensional Profiling of Micro-Nano Plastic in Bottled Water

Quantification from single-particle images with identified plastic polymer composition provides multi-dimensional information to build the analytical panorama of underexplored nanoplastics in bottled water.

Number quantification through particle counting suggests that on average, 78 to 103 plastic particles were identified in each FOV (0.2 mm × 0.2 mm) for three different brands, which was significantly higher (P < 0.001) than the blank samples (Fig. 5A). Assuming a uniform distribution of micro-nano plastic particles on the surface of the membrane region (SI Appendix, Supplementary Note 5), we can make an estimation for the micro-nano plastic exposure from bottled water. We estimate that there are on average about 2.4 ± 1.3  10⁵ plastic particles ingested from every liter of bottled water measured from different brands(Fig. 5C). Individual particles of each type of polymer are analyzed separately to reveal chemical heterogeneity. Within the library, PA, PP, PET, PVC, and PS are found likely to play a significant role in micro-nano plastics exposure from bottled water (Fig. 5B). The exact chemical composition of the micro-nano plastics varied from brand to brand, but PA seem to be the common major contributors in number among all the three brands we analyzed.

Harnessing the linear relationship between SRS intensity and the amount of analytes within the focal volume, we are also able to provide an estimation of exposure in mass besides particle number. The mass calibration curve can be estimated for each polymer out of density and relative SRS intensity from the linear relationship obtained by standard PS nanospheres (SI Appendix, Fig. S16). Integrated intensity within the region of interest for each particle is thus converted to mass (Fig. 5 E and F). The estimated micro-nano plastic exposure in mass is calculated to be at the level of around 10 ng/L. Analyzing the chemical composition in mass, we find unneglectable differences between contribution quantified by mass and contribution by number. Take the results from Brand C as an example. The PS nanoplastics though dominated in particle number, only account for a minor portion of the mass. Instead, PET becomes the major contributor together in mass. Such seeming disparity highlights the potential misunderstanding of plastic composition from collective particle characterization, which originated from the heterogeneous nature of micro-nano plastics from real-world samples.

Morphological characterization of individual particles enabled by SRS microscopy directly reveals another dimension of particle heterogeneity. Statistical analysis of particle size and shape from the images of individual micro-nano particles with well-defined identities is reported. When measuring the size distribution, we are able to characterize particles below the diffraction limit by extrapolating the size from the intensity reading (assuming the particles as solid spheres) and by using the linear relationship between the volume of the particles and SRS signal as calibration (SI Appendix, Supplementary Note 3). As a result, we find that plastic particles of different chemical compositions actually have different size distribution patterns (Fig. 6 A–G). The direct observation of the particle heterogeneity here provides a natural explanation of chemical compositional differences observed from mass or number measurement. Take PS and PET as an example: the size distribution of PS particles centers around 100 to 200 nm, whereas PET particles tend to have a size distribution that nears 1 to 2 microns, which explains why PET is a more significant component when measuring in mass while PS clearly dominates when counting the number of particles (Fig. 5 D and F).

The shape is another important morphological feature that matters as a critical aspect of nanotoxicity. Studies have shown that shape plays a role in determining the cellular uptake of micro-nano particles (65, 66). SRS images of plastic particles confirmed the existence of shape diversity for micro-nano plastics in bottled water. To account for the shape of plastic particles in a statistical manner, we measure the aspect ratio of individual particles above the diffraction limit (Fig. 6H). The aspect ratio is widely acknowledged in nanotoxicology studies (67, 68). The aspect ratio of the plastic particles detected ranges from 1 to 6, and the average aspect ratio for particles is around 1.7. Fig. 6 I–M provides a pictorial view of how the aspect ratio is related to the particle shape. Particles with an aspect ratio of above 3 are most likely to be fibrous in shape, while particles with an aspect ratio of below 1.4 will be largely spherical. Shape variation on plastic particles has been found in all polymers detected, confirming the widely recognized idea that real-world micro-nano plastics have diverse morphological prosperities. This dimension is hard to be resembled by engineered polymer nanoparticles commonly studied in research laboratories, and the toxicological consequences pertaining to real-life plastic particle exposures and their differing physicochemical properties (i.e., size, shape) have yet to be determined.

6. Discussions and Conclusions

By developing the data-driven hyperspectral SRS imaging platform for micro-nano plastic analysis, we describe a methodology to improve nanoparticle detection sensitivity and polymer identification specificity, which has allowed us to start to address the long-lasting knowledge gap of nanoplastics. We estimate that the exposure to the micro-nano plastics from regular bottled water was at the level of 10⁵ particles per liter, which is two to three orders of magnitude more than the previously reported results merely focusing on large microplastics (SI Appendix, Table S4) (58, 59, 61, 69, 70). As it pertains to the estimation of human exposure, these values are substantially higher than those currently reported in the literature (56, 71), which is a result from the newly detected nanoplastic fraction of plastic particulate. The tiny particles previously invisible under conventional imaging actually dominate in number and account for ~90% of the entire population of plastic particles detected. The remaining 10% identified as microplastics have a concentration of around 3 × 10⁴ particles per liter (SI Appendix, Fig. S17), with the majority of them in the size below 2 µm. Larger particles (>2 µm), which are easier to identify under regular optical microscopy, are in the same order of magnitude as the reported microplastic analysis depending on the detection limited reported based on different technologies (SI Appendix, Fig. S17 and Table S4). Our results confirm the plastic fragmentation beyond the micron level by unambiguously detecting nanoplastics in real-life samples. Similar to many other particle size distributions in the natural world, there are substantially more nanoplastics, despite being invisible or unidentified under conventional particle imaging techniques, than previously counted large micron ones. This population of nanoplastics can be easily overlooked in mass quantification as well since nanoparticles with smaller sizes contain cubic-less substances. However, given the capability of these nanoplastic particles to cross the biological barrier, nanoparticles, despite the seemingly trivial contribution to the mass measurement, might play a predominant role in terms of toxicity evaluation (72, 73).

We also find many detected particles present SRS spectra that do not match any of the standards. In fact, our small library of seven plastic polymers can only account for roughly about 10% of the total particles/dots imaged under SRS microscopy. A similar level of identification rate is reported in the microplastic analysis in bottled water using vibrational microscopy, indicating the complicated particle composition inside the seemingly simple water sample (SI Appendix, Table S4). In this sense, if we assume all detected organic particles originate from plastics [the same assumption entailed by the quantitative result from SEM-EDX or Nile Red staining (19, 74)], the micro-nano plastic concentration could be as high as 10⁶ particles per liter. However, the common existence of natural organic matter certainly requires prudent distinction from spectroscopy with polymer specificity. Moreover, careful investigation of unidentified particles suggests other aspects that further increase the complexity of identifying chemical composition. For example, some particles exhibit identical features to the characteristic two peaks (C=O ester bond: 1,730 cm⁻¹; C=C double bond: 1,615 cm⁻¹) of the PET in the fingerprint region but present a great variety of vibrational peaks in the high-frequency C–H region (SI Appendix, Fig. S8 a–d). It is unlikely for a polymer material distinct from PET to display both the C=O and C=C vibrational signatures that perfectly match the standard PET spectrum. A more plausible explanation is that they are small heteroaggregates containing PET and other components, with their SRS spectrum being the superposition of the spectrum from each component. Indeed, for some larger ones, we can even capture the spatial chemical heterogeneity within the aggregates (SI Appendix, Fig. S8 a, e, and i). The possible formation of heteroaggregates between nanoplastics or other natural organic matter has long been recognized as a potential challenge in the analysis of nanoplastics and may influence toxicological outcomes within a biological exposure (11). Direct visualization of such heteroaggregates here in real-world samples supports such concerns. For other possible heteroaggregates formed without PET, rigorous identification will require expanding the spectral library and advancing analytical algorithms for SRS microscopy or other vibrational imaging techniques with extended spectral windows to address challenges imposed by massive particle heterogeneity (27, 75, 76).

Another important insight is that the particle size distribution varies with the different chemical compositions, suggesting an interconnection between particle morphology and chemical composition. The observed nonorthogonality between plastic composition and particle morphologies challenges the conventional assumption for micro-nano plastics characterization from ensemble measurement. Take the result from brand C analysis as an example, ensemble measurement of micro-nano plastics might suggest that the major substance is PET from compositional analysis and most of the plastic particles have sizes below 500 nm from the morphological analysis. Assuming the two dimensions as being independent properties, people might have an impression that most of the plastic particles in the bottled water from brand C should be PET particles with a size below 500 nm. However, our result from single-particle analysis presents a clear disparity: the sample turns out to contain a small number of PET particles of about micron size and a large number of PS particles with size below 500 nm.

Such nonorthogonality might provide valuable information to understand, trace, and eventually prevent possible sources of micro-nano plastic contamination. Specifically in drinking water production, plastic contamination is confirmed in every step from the well to the bottle (77). The discovered size differences among different plastic polymers might indicate precious information about contamination sources during water production. For example, PET and PE, which are used as the packaging material for bottled water for all three brands we analyzed, have similar size distribution patterns, with a major population of micron sizes compared to other polymers. A possible explanation is that some particles of this kind are newly released from the bottle package during transportation or storage, which are retained faithfully in the water sample. Other polymers such as PA, PP, PS, and PVC, which are not the packaging material but also identified with significant numbers, are most likely introduced before or during water production. PP and PA, which share the same broad distribution of sizes, are widely used as equipment components or coagulant aids in water treatment (78). Particularly, PA is the most popular membrane material used in reverse osmosis (79), which is a common water purification method shared by all three brands. PVC and PS, which have a unique size distribution favoring small nanoplastics, might indicate a contamination source even earlier. PVC is identified to be the most abundant polymer type in raw water from microplastic analysis (77). PS is known to be used as backbone material for ion exchange resins in water purification (80). It is possible large particles of PVC or PS get removed by the RO membranes in the later step of the water treatment, leaving mostly nano populations.

Lastly, the interconnection between particle morphology and chemical composition has profound implications for toxicological concerns. As studies with engineered nanoparticles have suggested and investigations of plastic particles are starting to indicate, toxicity induced by micro-nano particles is not only dose-dependent but also related to particle physicochemical characteristics and their effect on cellular interactions and uptake (81, 82). In the case of bottled water from brand C, the cytotoxicity induced by PS nanoplastics plus a small number of PET microplastics would be presumably different from the effect assumed from PET nanoparticles. True comprehensive toxicity evaluation for micro-nano plastics would require multidimensional characterization of plastic particles and the integration of each individual plastic particle regarding their divergent properties on chemical composition and particle morphologies. Single-particle imaging with nanoparticle sensitivity and plastic specificity provides indispensable information to address the rising toxicity concern. Not only it enables plastic particle profiling with accurate exposure quantification, but also it has a unique potential to directly visualize the particle-biology interactions. Therefore, we envision that the data-driven hyperspectral SRS imaging platform will continue bridging the gap of knowledge on plastic pollution at the nano level with an expanded spectral library to study more complicated biological and environmental samples.

7. Materials and Methods