PaperlayerView exact revision on arXiv

Source-linked AI summary

Interpretable and Explainable Machine Learning for Materials Science and Chemistry

Felipe Oviedo, Juan Lavista Ferres, Tonio Buonassisi, Keith Butler

arXiv:2111.01037v2cond-mat.mtrl-scics.LG

TL;DR

Materials-science machine learning needs explainability beyond prediction so experts can inspect models and extract scientific knowledge. This paper surveys interpretability and explainability methods across applications, then discusses challenges and emerging approaches, emphasizing that model interpretation alone does not establish causation or generalization.

  • Problem

    Materials-science machine learning needs explainability beyond prediction to support expert assessment of model issues, trust, and scientific insight.

  • Method

    The paper synthesizes interpretability and explainability methods and their applications across experiments, simulations, materials development, and complex systems.

  • Results

    The account presents interpretable ML as a tool for understanding predictions, guiding experiments, and extracting or expanding knowledge from materials data.

  • Takeaways & Limitations

    Interpretability methods can reveal relationships, chemical rules, important features, and physically relevant latent representations across materials and chemistry studies.

  • Takeaways & Limitations

    Model interpretation alone does not establish causation or generalization and should be supplemented by suitable experiments, causal modeling, and uncertainty estimates.

Abstract

from arXiv · show

While the uptake of data-driven approaches for materials science and chemistry is at an exciting, early stage, to realise the true potential of machine learning models for successful scientific discovery, they must have qualities beyond purely predictive power. The predictions and inner workings of models should provide a certain degree of explainability by human experts, permitting the identification of potential model issues or limitations, building trust on model predictions and unveiling unexpected correlations that may lead to scientific insights. In this work, we summarize applications of interpretability and explainability techniques for materials science and chemistry and discuss how these techniques can improve the outcome of scientific studies. We discuss various challenges for interpretable machine learning in materials science and, more broadly, in scientific settings. In particular, we emphasize the risks of inferring causation or reaching generalization by purely interpreting machine learning models and the need of uncertainty estimates for model explanations. Finally, we showcase a number of exciting developments in other fields that could benefit interpretability in material science and chemistry problems.

CONSPECTUS

This account surveys interpretable and explainable machine learning applications across materials science and chemistry, emphasizing scientific insight beyond predictive performance. It also identifies challenges, including risks around causation and generalization and the need for uncertainty estimates.

  • Explainability can help experts identify model issues, build trust in predictions, and uncover unexpected correlations relevant to scientific discovery.
  • The paper presents examples spanning first-principles calculations, physicochemical characterization, materials development, and integration into complex systems.
  • The authors emphasize that interpreting machine-learning models alone can risk inferring causation or reaching unwarranted generalizations.
  • The account highlights uncertainty estimates for model explanations and developments from other fields as priorities for materials-science applications.

INTRODUCTION

The introduction frames interpretability as a response to concerns that machine learning can lose scientific knowledge, then surveys methods and applications across materials disciplines. It presents interpretable ML as a route to extracting knowledge while acknowledging field-specific challenges.

  • The paper asks whether scientific knowledge is inevitably lost in machine-learning studies and how it can instead be extracted for scientific research.
  • It draws on prior taxonomies while providing a broad account of interpretability and explainability across materials-science disciplines.
  • The review covers inherently interpretable models and techniques that extract explanations from models.
  • Many highlighted methods can be implemented with existing software, while frontier approaches may reveal physical insights from materials data.
  • The introduction also discusses roadblocks and challenges specific to machine-learning interpretability in materials science.

KEY CONCEPTS

The paper begins by clarifying interpretability terminology because the field contains substantial technical jargon. It distinguishes classical and deep machine learning as a foundational terminology choice.

  • The account clarifies unavoidable interpretability terminology to make the discussion accessible and avoid ambiguity.
  • Classical machine learning denotes methods that are not neural-network-based, whereas deep learning denotes methods based on neural-network architectures.
  • The distinction matters because classical and deep learning methods generally require different interpretation approaches.

A b X

The key concepts distinguish intrinsic from extrinsic and local from global interpretations, while framing explanations as trade-offs among interpretability, completeness, and scientific correctness. The paper cautions that model explanations can suggest hypotheses without establishing physical causation.

  • Scientific explanations add correctness, meaning the degree to which an explanation is scientifically accurate from physical or chemical perspectives.
  • Interpretability concerns human understanding, while completeness concerns how accurately an explanation reflects the model’s actual operation.
  • Model explanations often trade completeness against interpretability, as complete mathematical descriptions can be less understandable than partial attribution maps.
  • Interpreting a model may indicate predictive variable importance or rough cause-and-effect ideas, but causation requires controlled experiments or explicit causal modeling.
  • Global explanations describe general model behavior, whereas local explanations explain a decision for a particular case or data point.
  • Intrinsic interpretation examines the model itself, while extrinsic interpretation examines how the model responds to inputs or other stimuli.

INTRINSICALLY INTERPRETABLE MODELS

Materials models can use physical descriptors and interpretable model structures to connect composition or structure with properties, but simpler models can misrepresent nonlinear relationships and feature importance can be misleading.

  • INTRINSICALLY INTERPRETABLE MODELS: Atomistic materials models commonly represent composition and structure through physically motivated descriptors or learned representations.Descriptor-based models encode environments and symmetries, while neural networks can learn representations directly from raw structural inputs.
  • INTRINSICALLY INTERPRETABLE MODELS: Descriptor vectors built from elemental-property statistics can be fitted with linear, tree-based, or ensemble models to reproduce target properties.Examples include means, sums, and standard deviations of electronegativity or valence-electron properties.
  • INTRINSICALLY INTERPRETABLE MODELS: Linear models expose feature contributions through fitted coefficients, while tree-based models support interpretation through their structure.Coefficient confidence bounds can also inform interpretation of linear and generalized linear models.
  • INTRINSICALLY INTERPRETABLE MODELS: Feature importance scores can favour continuous over categorical features, especially when categorical features are high-dimensional or continuous features have wide ranges.These package-level pathologies require caution when interpreting feature importance in materials datasets.
  • INTRINSICALLY INTERPRETABLE MODELS: Regularisation methods such as SISSO and LASSO reduce descriptor dimensionality, but correlated features can make selected features arbitrary.Correlation metrics can help assess feature-elimination results, although low correlation does not establish that features are unrelated.
  • INTRINSICALLY INTERPRETABLE MODELS: Model mismatch can make simple models assign importance to irrelevant features, whereas GAMs can represent nonlinear relationships with high interpretability.The paper illustrates this trade-off using chemical adsorption on subsurface alloys.

MODEL INTERPRETATION METHODS

Complex models often require extrinsic interpretation methods, which can supplement intrinsic feature examination with insights unavailable from the model alone.

  • MODEL INTERPRETATION METHODS: Extrinsic interpretation methods are useful for complex models such as deep neural networks and can add insights even when feature importance is already interpretable.The paper introduces model-agnostic explanations before methods designed specifically for deep neural networks.

What-if interpretations

What-if methods explain predictions by changing inputs and observing outputs; SHAP can reveal descriptor relationships, but its attributions do not establish causal structure and may use unrealistic combinations.

  • What-if interpretations: Partial dependence plots estimate how changing one feature affects the output while marginalising over other features.The paper gives mean atomic mass and dielectric response as a materials example.
  • What-if interpretations: SHAP analysis revealed descriptor relationships for crystal dielectric constants that resembled established empirical models while providing greater predictive power.The approach was applied to a support vector regression model.
  • What-if interpretations: SHAP values can sample physically unrealistic combinations when correlated parameters are varied, and they do not represent direct causal relations between features.The paper gives HOMO energy, LUMO energy, and band gap as an example of these limitations.

Deep interpretations

Deep-learning interpretations use salience or representation methods to identify influential input regions, including chemically or physically meaningful patterns, but these explanations can be over-interpreted.

  • Deep interpretations: Deep-learning interpretation methods examine either how a model processes an input or the representations it learns.The paper classifies these as processing methods and representation methods.
  • Deep interpretations: Salience methods identify input regions associated with classification by testing sensitivity or activation patterns.Class activation maps project weighted final-layer filters back onto the input to show important regions.
  • Deep interpretations: Transformer attention mechanisms learn weighted relationships among input elements, and aggregated attention masks can identify task-relevant regions.The paper describes applications to molecular fragments, chemical reaction atom-mapping, and materials bandgap prediction.
  • Deep interpretations: CAM identified energy-transfer regions in Q-space for classifying magnetic Hamiltonians that matched regions identified by a trained physicist, in a fraction of the time.The comparison concerns inelastic neutron-scattering spectra classified by a convolutional neural network.
  • Deep interpretations: Attention maps identified molecular reaction sites, chemical grammars, and n-type dopants receiving attention in bandgap prediction for silicon-containing materials.These findings are presented as examples of chemically or physically meaningful correlations extracted from learned attention.
  • Deep interpretations: Salience and attention maps can support over-interpretation because identical maps may occur across classes or models may exploit correlations without physical causes.The paper advises combining interpretations with secondary experimentation or simulation.

EXPERIMENTAL PREDICTIONS AND EXPLANATIONS

Interpretability supports experimental characterization, optimization, sensitivity analysis, and hypothesis testing by connecting model behavior with physical parameters and experimental decisions. Intrinsic and extrinsic approaches can improve trust and strengthen the relationship between algorithms and humans in the experimental loop.

  • Experimental model use: Interpretability facilitates characterization, optimization, sensitivity analysis, and hypothesis testing in experimental materials research.Models can identify or optimize dominant physical parameters, although explanations remain subject to tensions among correctness, interpretability, and completeness.
  • Experimental model use: Intrinsic physical parametrization and surrogate ML models can map fabrication variables to figures of merit in layered semiconductors.First-principles calculations can also constrain surrogate-based compositional optimization.
  • Experimental model use: Predictive power may be secondary to induction or explainability when experiments test hypotheses, especially where models identify dominant descriptors or guide physically constrained interventions.Likelihood-free inference can help represent complicated physical and chemical systems when high-performing physical simulations are limited.
  • Extrinsic explanations: SHAP and salience methods connect ML models to experimental procedures by identifying influential molecular features or informative regions in spectroscopy signals.SHAP highlighted low numbers of hydrogen bond donors and small topological polar surface areas for organic capping layers in halide perovskite solar cells.
  • Practical significance: Interpretability increases trust in ML approaches and strengthens the relationship between algorithms and humans in the experimental loop.

PHYSICAL KNOWLEDGE BEYOND MODEL EXPLANATIONS

Beyond explaining existing models, interpretable ML can directly extract physical knowledge from data through sparse models, genetic algorithms, and constrained latent representations. However, explanations still lack reliable uncertainty and causal guarantees, motivating confirmation through experiments, simulations, or causal methods.

  • Knowledge extraction: Scientific ML problems may be better framed around extracting knowledge or causal interpretations rather than treating interpretability as a secondary predictive goal.The authors identify this framing as a potentially better trade-off in exploratory research.
  • Knowledge extraction: Sparse regression, genetic algorithms, and deep autoencoders provide progressively broader strategies for finding functional forms or representations that explain noisy physical data.Autoencoders learn compressed latent representations by reconstructing inputs through a reduced-dimensional space.
  • Knowledge extraction: Latent spaces can be constrained with physical penalties, hierarchical or causal graphical structures, adversarial training, and related representation-design methods.These constraints aim to make learned representations more physically meaningful or interpretable.
  • Knowledge extraction: β-VAE methods impose orthogonality and sparsity so latent dimensions are uncorrelated and only the minimum number needed for reconstruction is used.Such representations have been shown to extract parameters interpretable as driving parameters of ordinary differential equations from dynamic-process data.
  • Challenges: Model explanations commonly lack confidence intervals or error distributions, limiting confidence in insights extracted from modern machine-learning models.The paper points to work addressing uncertainty and bias in explanations and expects future approaches to integrate uncertainty into extracted insights.
  • Challenges: Interpretability explanations lack strong causal guarantees and resilience against confounding, so real-world insights require scientist judgment or secondary experimental or simulation confirmation.The paper distinguishes correlations identified by most ML methods from cause-and-effect relationships.
  • Causal inference: Combining ML with causal inference has been used to uncover mechanisms driving ferroelectric distortions from electron microscopy data.Traditional experimental design can provide substantial control of confounding factors in hard experimental science.
  • Causal inference: Shapley explanations can sample unrealistic parameter combinations and do not encode causality, while VAEs and asymmetric Shapley formulations are proposed to address these issues.The cited developments had not yet been applied to materials-science ML in the discussed account.

CONCLUSIONS

The paper surveys interpretable ML methods across materials science and argues that understanding models and extracting knowledge should accompany their growing use. It presents interpretability as an initial approximation to causal hypothesis generation, not a replacement for human expertise.

  • Conclusions: The paper outlines interpretable ML methods spanning the range of approaches increasingly used in materials science.
  • Conclusions: Direct knowledge-extraction approaches include sparse regression or regularization, chemical or physical parameters or functions, and latent-space representations with convolutional operators.Figure 5 presents a taxonomy and a solar-cell autoencoder example in which current-voltage characteristics are encoded into physical parameters for operational knowledge and device-optimization paths.
  • Conclusions: The authors hope future materials-science ML papers will include efforts to understand derived models and extract more knowledge from data.They note that many presented interpretability methods are as easily implemented as the ML models they interpret.
  • Conclusions: Interpretable ML is not a silver bullet for model understanding and constitutes only an initial approximation to causal hypothesis generation.
  • Conclusions: Human experts remain central to using interpretable ML in materials science.
Loading 2111.01037v2…