spectral feature attribution

This work addresses the limitations of existing explainable AI methods in interpreting spectral machine learning models, which often overlook chemically meaningful contiguous spectral regions. To bridge this gap, the authors propose SMX, a novel framework that incorporates chemical prior knowledge into model interpretation by leveraging expert-defined spectral intervals to deliver global, post-hoc, and model-agnostic explanations. SMX integrates principal component analysis (PCA), logical predicates, and perturbation analysis, and employs a directed weighted graph to aggregate region-wise importance scores. Furthermore, it introduces a threshold-based spectral reconstruction technique to intuitively visualize explanations directly in the original spectral domain. Experimental results across eight real-world datasets—comprising six X-ray fluorescence (XRF) and two gamma-ray spectra—as well as one synthetic benchmark demonstrate that SMX significantly outperforms current approaches in both explanatory fidelity and interpretability.

This study addresses the instability in model interpretability caused by high-dimensional and highly collinear spectral data, which often obscures the relationship between original signals and predictive rationale. To resolve this, the authors propose a novel approach that integrates principal component analysis (PCA) with SHAP (SHapley Additive exPlanations). The method first employs PCA for dimensionality reduction to enhance model stability, then maps SHAP-based explanations back to the original spectral space, thereby preserving both global and local interpretability. This work is the first to achieve consistent representation of feature importance—derived from reduced dimensions—in the original input space, effectively mitigating the interpretability disconnect commonly induced by conventional dimensionality reduction. Experimental results demonstrate that the proposed method significantly improves explanation stability across repeated runs and accurately links critical spectral bands to their corresponding biochemical components, thereby enhancing model trustworthiness and practical utility.

This study addresses the contradictory conclusions in convolutional neural network (CNN) design for visible–near-infrared (Vis-NIR) chemometrics—such as inconsistent findings regarding kernel size, network depth, and the efficacy of preprocessing or transfer learning—by identifying uncontrolled moderating variables as the root cause. To resolve this, the authors propose a conditional design framework that explicitly links model architecture and preprocessing strategies to the physical properties of spectra, dataset characteristics, and deployment scenarios. For the first time, this framework integrates effective receptive field analysis, the indirect measurement nature of water matrices, and validation protocol design. The resulting approach establishes interpretable and reproducible CNN design principles, significantly enhancing model stability and generalization across diverse datasets and real-world deployment environments.

To address the low accuracy of automated infrared (IR) spectral interpretation under data-scarce conditions, this paper proposes the first large language model (LLM)-driven agent framework tailored for IR spectroscopy analysis. The framework integrates automated spectral preprocessing, literature-based knowledge retrieval, few-shot prompting, and a multi-round closed-loop reasoning mechanism to enable end-to-end multi-task analysis—including functional group identification, material classification, and compositional inference. Its key contributions are: (1) the first application of LLM-based agent systems to IR spectral analysis; (2) a dynamic feedback-enabled multi-round reasoning strategy that substantially improves few-shot generalization; and (3) cross-material applicability—demonstrated on diverse domains such as seal ink, traditional Chinese medicine, and Pu’er tea. Experiments show that, in low-data regimes, our method matches or surpasses conventional machine learning and deep learning baselines, with multi-round reasoning yielding significant performance gains over single-round inference.

This study addresses the lack of generality and interpretability in applicability domain (AD) estimation for machine learning models. We propose a unified AD assessment framework based on kernel density estimation (KDE), which quantifies the distance of a query sample from the training data distribution in feature space and establishes a quantitative relationship among distance, prediction error, and uncertainty. Chemical prior knowledge is incorporated to calibrate the AD decision threshold. To our knowledge, this is the first method enabling consistent, cross-model and cross-task AD evaluation across diverse models—including random forests (RF), gradient-boosted decision trees (GBDT), and graph neural networks (GNN)—and heterogeneous materials datasets (crystals, molecules, alloys). Experiments demonstrate that large KDE-derived distances strongly correlate with high prediction residuals and elevated uncertainty estimates. An open-source toolkit enables automated in-domain/out-of-domain classification. The implementation and documentation are publicly available.

This study addresses the prevalent use of generic convolutional neural network (CNN) architectures in near-infrared (NIR) spectroscopic chemometrics, which often overlooks critical data-specific characteristics such as sampling properties, smoothness, redundancy, and sample size. By systematically analyzing 25 NIR regression tasks, the authors develop an interpretable one-dimensional CNN backbone and integrate spectral descriptors—including spectral entropy, intrinsic rank, autocorrelation, and wavelet scales—to establish, for the first time, a mapping between spectral properties and optimal CNN hyperparameters. This mapping is derived via Bayesian hyperparameter optimization and leave-one-dataset-out validation. The resulting transferable warm-start heuristic significantly reduces reliance on costly hyperparameter searches, achieving median test RMSE ratios of 0.953 and 1.017 under direct and leave-one-out evaluation, respectively—performance comparable to fully optimized models with equivalent sensitivity to random seeds.

This study addresses the challenge of efficiently and accurately predicting soil properties from high-dimensional, highly collinear spectral data across multiple spatial scales—from field to global. We systematically evaluate the performance of various regression models and dimensionality reduction techniques across 85 prediction tasks. Notably, we introduce TabPFN, a tabular foundation model, into soil spectroscopy for the first time, demonstrating its ability to outperform conventional approaches without explicit dimensionality reduction. Further gains in predictive accuracy are achieved by integrating partial least squares (PLS) latent variables with TabPFN. Empirical results show that TabPFN consistently achieves state-of-the-art performance across all scales, substantially surpassing classical baselines—particularly on large-scale global spectral libraries comprising tens of thousands of samples—thereby highlighting its potential as a general-purpose modeling paradigm for soil property prediction.

This study addresses the critical challenge of wavelength selection in near-infrared spectroscopy for accurate and interpretable sugar content prediction, a task often hindered by noise sensitivity and instability in existing methods. To overcome these limitations, the authors propose the first application of combinatorial Bayesian optimization to this domain, formulating wavelength region selection as a binary black-box optimization problem. They construct a sparse quadratic surrogate model and solve the resulting quadratic unconstrained binary optimization (QUBO) problem using Thompson sampling combined with classical or quantum annealing. Compared to genetic algorithms and conventional simulated annealing, the proposed approach achieves significantly higher prediction accuracy. Moreover, the selected wavelength bands exhibit minimal root-mean-square error variation under single-bit perturbations, demonstrating superior consistency, robustness, and generalization capability.

本文提出S3C-LLM，通过技能指导和代码执行的方式解决光谱到结构解析的问题，改进了现有基于大语言模型的方法，实现了更准确的分子结构预测。

本文通过影响函数解决ESA的Ariel任务中机器学习模型的可解释性问题，提出了一种基于预测的影响函数方法，并利用该方法计算误差代理。