Interpretable machine learning as a tool for scientific discovery in chemistry

Abstract

There has been an upsurge of interest in applying machine learning to chemistry, and impressive predictive accuracies have been achieved, but this has been done without providing any insight into what has been learnt from the training data.

Impact of the characteristics of quantum chemical databases on machine learning predictions of tautomerization energies

“…The prominence of ML has raised concerns regarding the 'interpretability' of the models conceived. 24,75 This awareness also increases for complex models because a rational relationship between initial data used for training and resulting prediction becomes less transparent. Therefore, it is important to develop quantifiable and intuitive tests for how ML models "work" and how trustworthy predictions by them are.…”

“…21,22 This process has also been called "interpretability" and it can be used to understand the relationships learned by the model or contained in the data used for training it. 23,24 Part of the present work is concerned with the aim to relate the composition of the initial chemical databases based on which ML models are conceived with their performance on the prediction of a property of interest (tautomerization energy) on a set of unseen examples.…”

Data-driven approach towards identifying dyesensitizer molecules for higher power conversion efficiency in solar cells

“…The innate cognitive attribute of any ML algorithm is to learn and advance into a better system over time when exposed to a new set of data. 14 In recent times, numerous ML techniques have been deployed in addressing real-time problems just by adopting regression classification-based approaches. There have also been some important ML applications in chemistry, such as drug discovery, 15,16 organic synthesis, 17 catalytic reactions, 18 materials design for batteries, 19 supercapacitors, 20,21 and fabrication of solar cells.…”

Explaining structure-activity relationships using locally faithful surrogate models

“…Explainable artificial intelligence (XAI) has emerged as a field to better comprehend what DL models learn and to gain scientific insight into model predictions. [23,24] Two broad approaches are typically used for model interpretability -intrinsic interpretability and post-hoc approaches. [25] Intrinsic interpretability comes from using models that are considered inherently interpretable or self-explaining.…”