Probing out-of-distribution generalization in machine learning for materials

Abstract: Scientific ML endeavors to develop generalizable models with broad applicability. However, the assessment of generalizability is often based on heuristics. Here, we demonstrate in the materials science setting that heuristics based evaluations lead to substantially biased conclusions of ML generalizability and benefits of neural scaling. We evaluate generalization performance in over 700 out-of-distribution tasks that features new chemistry or structural symmetry not present in the training data. Surprisingly, good performance is found in most tasks and across various ML models including simple boosted trees. Analysis of the materials representation space reveals that most tasks contain test data that lie in regions well covered by training data, while poorly-performing tasks contain mainly test data outside the training domain. For the latter case, increasing training set size or training time has marginal or even adverse effects on the generalization performance, contrary to what the neural scaling paradigm assumes. Our findings show that most heuristically-defined out-of-distribution tests are not genuinely difficult and evaluate only the ability to interpolate. Evaluating on such tasks rather than the truly challenging ones can lead to an overestimation of generalizability and benefits of scaling.

• The paper demonstrates that ML models robustly generalize across 700+ out-of-distribution material tasks, even with simple methods like boosted trees.
• The study reveals that many heuristic OOD tests primarily assess interpolation rather than true extrapolation to novel material domains.
• The analysis of representation space challenges neural scaling laws by showing that additional training may not enhance performance on genuinely new tasks.

Probing Out-of-Distribution Generalization in Machine Learning for Materials

The paper by Kangming Li et al. investigates the generalization capabilities of ML models when applied to out-of-distribution (OOD) tasks in the domain of materials science. Conventional evaluations of ML model generalizability often rely on heuristic methods, which the authors argue may lead to biased conclusions, particularly in the context of neural scaling benefits. In their study, the authors critically assess the performance of various ML models across over 700 OOD tasks that introduce new chemical or structural characteristics absent from training data.

A notable finding of this study is the generally strong performance of these ML models across a broad range of OOD tasks, even for relatively simple models like boosted trees. The analysis of the material representation space revealed that a substantial portion of the test tasks lies within regions already well-covered by training data. This suggests that many OOD tests primarily challenge models' interpolation capabilities rather than their ability to generalize to entirely new domains. Such findings challenge the prevailing assumption that increasing training set size or time, a key tenet of the neural scaling paradigm, will inherently enhance generalization capabilities. Instead, for tasks lying outside the training domain, additional training often produced little to no improvement or even degraded performance, contradicting traditional scaling principles.

Key Findings

1. Strong Generalization Across Chemistries: The study shows that current ML models exhibit robust OOD generalization to unfamiliar chemistries. This capability is evident across both low-complexity models, like random forests, and more sophisticated architectures, such as graph neural networks and LLMs. The results suggest that good OOD performance is frequently achievable, challenging the view that ML models struggle with chemical dissimilarity.
2. Challenge of Heuristic OOD Tasks: The authors demonstrate that many heuristically-defined tests are easy for existing models and mostly require interpolation rather than true OOD generalization. The study emphasizes the potential for misinterpreting certain OOD test results as emergent abilities of a model.
3. Insight into Representational Space: By examining the representation space, the paper identifies a distinction between representationally in-domain and out-of-domain test samples. This method efficiently uncovers the identifiable regions in representation space and provides insights into areas where ML models tend to fail due to lack of training data coverage.
4. Limitations of Neural Scaling Laws: The findings reveal that scaling training set sizes or computation does not substantially improve generalization to representationally out-of-domain tasks, shedding light on the overestimation of neural scaling benefits.

Implications and Future Directions

This work has significant implications for the development of materials science ML models. By highlighting the tendency of heuristic-based OOD tasks to overestimate model generalizability, it encourages the creation of more genuinely challenging benchmarks. This will aid in thoroughly evaluating ML models' ability to extrapolate beyond the training distribution. Moreover, the findings suggest that future ML advancements in materials science should target enhancing representation models and leveraging domain-specific knowledge to intelligently identify truly challenging generalization tasks. Additionally, improving feature representation quality and designing more sophisticated ML architectures may bridge the gap in generalizability without overly relying on scaling data or computational resources.

The study informs efforts to develop foundational models in scientific machine learning, particularly in materials discovery. The authors' insights into the complexity of domain identification highlight the necessity for a nuanced understanding of training and test distribution overlap, which is paramount for model development strategies aiming at robust and truly generalizable solutions.