Identifying domains of applicability of machine learning models for materials science

Abstract

Although machine learning (ML) models promise to substantially accelerate the discovery of novel materials, their performance is often still insufficient to draw reliable conclusions. Improved ML models are therefore actively researched, but their design is currently guided mainly by monitoring the average model test error. This can render different models indistinguishable although their performance differs substantially across materials, or it can make a model appear generally insufficient while it actually works well in specific sub-domains. Here, we present a method, based on subgroup discovery, for detecting domains of applicability (DA) of models within a materials class. The utility of this approach is demonstrated by analyzing three state-of-the-art ML models for predicting the formation energy of transparent conducting oxides. We find that, despite having a mutually indistinguishable and unsatisfactory average error, the models have DAs with distinctive features and notably improved performance.

Identifying Outstanding Transition-Metal-Alloy Heterogeneous Catalysts for the Oxygen Reduction and Evolution Reactions via Subgroup Discovery

“…We start our analysis by introducing the SGD approach [14][15][16][17][18] to uncover complex patterns associated to outstanding local behavior by using data sets. This methodology has been recently applied to catalysis [22] as well as materialsscience [18,23] problems.…”

Big data and machine learning for materials science

“…Furthermore, even in the state of the art of AI and ML, there are no clear hints as to whether and how uses will be possible within the “explanation” category, not only for materials science but also for other areas. These limitations have been discussed in a proposal to employ AI and uncertainty quantification to obtain correctable models [ 176 ] and in identifying domains where ML is applicable more efficiently [ 177 ]. One may speculate that the answer may result from the convergence of the two big movements mentioned in the Introduction – big data and natural language processing–but the specifics of the solutions are far from established.…”

Machine learning dielectric screening for the simulation of excited state properties of molecules and materials

“…Another important advantage of the ML-derived dielectric screening is that it provides insight into the approximate screening parameters used in the derivation of hybrid functionals for time-dependent DFT (TDDFT) calculations, including dielectric-dependent hybrid (DDH) functionals. [44][45][46][47][48] We emphasize that the strategy adopted here is different in spirit from strategies that use ML to infer structure-property relationships [49][50][51][52][53][54][55][56][57] or relationships between computational and experimental data. 58 We do not seek to relate structural properties of a molecule or a solid to its absorption spectrum.…”