Machine learning has revolutionized materials informatics, enabling rapid screening of high-entropy alloys, predicting crystal stability, and accelerating molecular dynamics via neural network interatomic potentials. However, a fundamental tension exists: the most accurate models are often deep, non-linear "black boxes."
As noted across the field, from high-entropy alloy discovery (Ruixuan Li et al. in High-Throughput Calculations for High-Entropy Alloys: A Brief Review) to intelligent fiber engineering (Yuanlong Shao et al. in Artificial intelligence reforges intelligent fibers), a model that predicts a property perfectly without explaining why provides limited scientific value. It fails to yield transferable chemical principles or actionable design rules (Chenru Duan et al. in Machine Learning for the Discovery, Design, and Engineering of Materials).
To bridge this explainability gap, researchers are developing methods to extract physical intuition from data-driven models. These approaches broadly fall into three paradigms: inherently interpretable "glass-box" architectures, post-hoc explanation mechanisms, and hybrid structure-property mapping.
Inherently Interpretable Glass-Box Models
Instead of training an opaque neural network and attempting to decode its weights post-hoc, one strategy is to constrain the model's architecture so that its internal mathematics map directly to physical concepts.
In the domain of interatomic potentials—where deep learning models often obscure how they identify local atomic configurations (Bangchen Yin et al. in AlphaNet: scaling up local-frame-based neural network interatomic potentials)—inherent interpretability prevents unphysical extrapolations. In Ultra-fast interpretable machine-learning potentials, Stephen Xie and Matthias Rupp developed potentials that combine a flexible B-spline basis with regularized linear regression. Because the model relies on linear combinations of explicit basis functions, the energy contributions of specific interatomic distances are directly inspectable, matching the transparency of traditional empirical potentials while capturing quantum-mechanical accuracy.
Similarly, for predicting thermodynamic stability, Justin Tahmassebpur et al. demonstrated in A seven-facet polyhedron captures the composition-only formation-energy landscape of inorganic solids that a complex composition-only formation-energy landscape can be captured by a simple, max-affine seven-facet polyhedron. This minimalist geometric representation allows researchers to extract reusable chemical principles directly from the boundaries of the polyhedron.
When explicit mathematical formulas are required for practical engineering, symbolic regression can bypass black boxes entirely. In Machine learning of mechanical properties of steels, Jie Xiong and San‐Qiang Shi used symbolic regression to discover closed-form analytical formulas for steel performance, isolating tempering temperature and key alloying elements as the primary physical drivers.
Post-Hoc Interpretability and Attention Mechanisms
When highly non-linear architectures like deep neural networks or gradient boosting trees are necessary to capture complex physics, researchers deploy post-hoc interpretability tools to peer inside the trained model.
One prominent mechanism is the integration of attention layers within neural networks. In the Self-Consistent Attention Neural Network (SCANN) framework developed by Tien-Sinh Vu et al. in Towards understanding structure–property relations in materials with interpretable deep learning, attention coefficients act as dynamic masks. By analyzing which coefficients fire during a prediction, the authors identified the exact atoms and local structural environments that drive properties like molecular orbital energies and crystal stability, transforming the network into a tool for causal discovery.
[Input Crystal Structure] ──> [Attention Layers (SCANN)] ──> [Property Prediction]
│
└───> High Attention Weights = Key Causal Atoms
For tabular datasets governing complex processing-structure-property relationships, game-theoretic approaches like SHAP (Shapley Additive exPlanations) are widely used. In An Interpretable Boosting-based Predictive Model for Transformation Temperatures of Shape Memory Alloys, Sina Hossein Zadeh et al. applied SHAP values to a gradient boosting model. This allowed the authors to isolate how specific chemical compositions and manufacturing parameters interact to dictate alloy behavior, transforming an opaque ensemble model into an intuitive guide for alloy synthesis.
Mapping and Clustering Structure-Property Spaces
A third paradigm involves blending unsupervised learning (which uncovers natural groupings in materials data) with supervised learning (which maps structures to properties). This helps scientists visualize the physical gradients within a dataset.
Benjamin A. Helfrecht et al. introduced this concept via Kernel Principal Covariates Regression (KPCovR) in Structure-Property Maps with Kernel Principal Covariates Regression. KPCovR optimizes a joint objective function that simultaneously minimizes unsupervised reconstruction error (like PCA) and supervised prediction error (like regression). The result is a low-dimensional "structure-property map" where spatial clusters correspond to structural similarities, and the spatial gradients correspond to changes in physical properties, making outliers and trends visually intuitive.
Unsupervised Loss (PCA-like) ──┐
├─> Joint Objective ─> Low-Dim Structure-Property Map
Supervised Loss (Regression) ──┘
In a similar vein, Vahe Gharakhanyan et al. combined unsupervised clustering with supervised regression in Discovering Melting Temperature Prediction Models of Inorganic Solids by Combining Supervised and Unsupervised Learning. By first grouping materials based on their atomic bonding characteristics, the authors uncovered that entirely different thermodynamic drivers (such as shear modulus versus cohesive energy) dominate the melting process depending on whether a material is ionic or covalent. A single global black-box model would obscure these distinct physical regimes.
This philosophy of using collective intelligence to find meaningful groupings is also seen in the committee machine approach by Duong‐Nguyen Nguyen et al. in Committee machine that votes for similarity between materials, where multiple regression models vote on material similarity to uncover underlying physical mechanisms.
Resolving the Dual Black-Box Problem
The push toward interpretability addresses what Jianpeng Chen et al. term the "dual black-box" problem in MetamatBench: Integrating Heterogeneous Data, Computational Tools, and Visual Interface for Metamaterial Discovery: materials scientists struggle to trust opaque machine learning outputs, while AI researchers lack the physical intuition to diagnose model failures.
By shifting the focus of materials informatics from pure predictive accuracy to physical interpretability, these diverse methodologies ensure that machine learning serves not just as an automated screening tool, but as a catalyst for fundamental scientific insight.
Key Foundational Work
Committee machine that votes for similarity between materials (2018) — contributed 1 concept elements to this direction.
Evolution in Materio: Exploiting the Physics of Materials for Computation (2008) — contributed 1 concept elements to this direction.