# Understanding and improving transferability in machine-learned activation energy predictors

Source: [Understanding and improving transferability in machine-learned activation energy predictors](<https://www.sota2.com/research/paper/68fa4a1e1e209131ce93bf4f>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-27T04:38:04.409Z

Authors: Joe Gilkes, Mark Storr, Reinhard J. Maurer, Scott Habershon

Published: 2025-05-01T15:35:12.000Z

## Abstract

The calculation of reactive properties is a challenging task in chemical reaction discovery. Machine learning (ML) methods play an important role in accelerating electronic structure predictions of activation energies and reaction enthalpies, and are a crucial ingredient to enable large-scale automated reaction network discovery with $\>10^3$ reactions. Unfortunately, the predictive accuracy of existing ML models does not yet reach the required accuracy across the space of possible chemical reactions to enable subsequent kinetic simulations that even qualitatively agree with experimental kinetics. Here, we comprehensively assess the underlying reasons for prediction failures within a selection of machine-learned models of reactivity. Models based on difference fingerprints between reactant and product structures lack transferability despite providing good in-distribution predictions. This results in a significant loss of information about the context and mechanism of chemical reactions. We propose a convolutional ML model that uses atom-centered quantum-chemical descriptors and approximate transition state information. Inclusion of the latter improves transferability for out-of-distribution benchmark reactions, making greater use of the limited chemical reaction space spanned by the training data. The model further delivers atom-level contributions to activation energies and reaction enthalpies that provide a useful interpretational tool for rationalizing reactivity.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2505.00604>)

[PDF](<https://arxiv.org/pdf/2505.00604.pdf>)
