Artificial intelligence has revolutionized material discovery by accelerating the screening of novel compounds, predicting properties, and optimizing synthesis pathways. However, the complexity of AI models often obscures the underlying physical mechanisms governing material behavior. To bridge this gap, interpretable AI methods are critical for extracting actionable insights. Three key approaches—SHAP values, attention mechanisms, and symbolic regression—have emerged as powerful tools for making AI-driven discoveries transparent and physically meaningful.

SHAP (SHapley Additive exPlanations) values quantify the contribution of each input feature to a model’s prediction, rooted in cooperative game theory. In material science, SHAP analysis has been applied to uncover non-intuitive relationships between dopants and electronic properties. For instance, a study on perovskite solar cells used SHAP values to reveal that certain halide substitutions disproportionately influenced defect tolerance, despite their low concentration. The analysis showed that bromine incorporation, even at minimal levels, significantly suppressed deep-level traps due to its impact on lattice strain. This insight guided targeted doping strategies, improving device efficiency without exhaustive trial-and-error experimentation. Similarly, in high-entropy alloys, SHAP values identified that local atomic distortions, rather than elemental diversity alone, dominated phase stability—a finding later validated by ab initio calculations.

Attention mechanisms in neural networks provide another layer of interpretability by highlighting which segments of input data the model prioritizes during decision-making. Transformer-based architectures, equipped with self-attention, have been employed to analyze crystal structures and predict properties such as bandgap and ionic conductivity. A notable case involved a study on solid-state electrolytes, where attention weights revealed that specific coordination polyhedra of lithium ions were critical for high ionic mobility. The model’s attention patterns aligned with known diffusion pathways, confirming the physical basis of its predictions. In another example, attention maps derived from a graph neural network trained on molecular semiconductors highlighted the role of side-chain stacking in charge transport, leading to the design of polymers with optimized backbone rigidity and side-chain spacing.

Symbolic regression goes a step further by distilling complex datasets into compact, interpretable equations. Unlike traditional regression, which imposes a fixed functional form, symbolic regression searches over a space of mathematical expressions to identify relationships that balance accuracy and simplicity. This approach has been instrumental in deriving physically meaningful laws from high-throughput computational or experimental data. For example, researchers applied symbolic regression to defect formation energies in transition metal oxides and uncovered a non-monotonic dependence on oxidation state, which conventional descriptors failed to capture. The resulting equation linked defect stability to ionic radii and electronegativity differences, enabling rapid screening of defect-tolerant materials. In thermoelectrics, symbolic regression identified a previously overlooked trade-off between lattice thermal conductivity and anharmonicity, expressed through a power-law relation involving Grüneisen parameters and atomic mass disparities.

The integration of these interpretable AI methods has also exposed unexpected material behaviors. A study on doped silicon carbide revealed, through SHAP analysis, that nitrogen vacancies played a dual role: while they acted as recombination centers, they also passivated carbon vacancies by charge compensation. This counterintuitive finding explained anomalous carrier lifetime measurements and informed co-doping strategies to mitigate recombination. Attention mechanisms, applied to X-ray diffraction data of battery cathodes, detected subtle peak shifts correlated with cation disorder, leading to the discovery of a metastable phase that enhanced lithium diffusion. Symbolic regression, when applied to the bandgap bowing in III-V alloys, derived a simple expression that accounted for both electronegativity mismatch and strain effects, resolving discrepancies between prior theoretical models.

Despite their strengths, these methods require careful implementation. SHAP values can be computationally expensive for high-dimensional datasets, necessitating dimensionality reduction or feature selection. Attention mechanisms may produce noisy interpretations if trained on insufficient or biased data, emphasizing the need for robust training sets. Symbolic regression faces challenges in balancing equation complexity with generalizability, often requiring constraints to avoid overfitting. However, when applied judiciously, these techniques transform AI from a black-box predictor into a collaborator that elucidates the physics behind material properties.

The future of interpretable AI in material discovery lies in hybrid approaches that combine these methods. For instance, attention-guided symbolic regression could focus equation discovery on the most relevant structural features, while SHAP values could validate the contributions of terms in the derived equations. Such integrations will be crucial for tackling grand challenges, such as predicting emergent phenomena in correlated electron systems or optimizing multi-component catalysts. By prioritizing interpretability, researchers can ensure that AI-driven insights are not just statistically sound but also scientifically illuminating, accelerating the translation of computational discoveries into real-world applications.