From Prescriptive to Predictive: an Interdisciplinary Perspective on the Future of Computational Chemistry

Abstract

Reliable predictions of the behaviour of chemical systems are essential across many industries, from nanoscale engineering over validation of advanced materials to nanotoxicity assessment in health and medicine. For the future we therefore envision a paradigm shift for the design of chemical simulations across all length scales from a prescriptive to a predictive and quantitative science. This paper presents an integrative perspective about the state-of-the-art of modelling in computational and theoretical chemistry with examples from data-and equation-based models. Extension to include reliable risk assessments and quality control are discussed. To specify and broaden the concept of chemical accuracy in the design cycle of reliable and robust molecular simulations the fields of computational chemistry, physics, mathematics, visualisation science, and engineering are bridged. Methods from electronic structure calculations serve as examples to explain how uncertainties arise: through assumed mechanisms in form of equations, model parameters, algorithms, and numerical implementations. We

Relative molecule self-attention transformer

“…The vanilla self-attention includes absolute encoding of position, which can hinder learning when the absolute position in the sentence is not informative. 1 Relative positional encoding featurizes the relative distance between each pair of tokens, which led to substantial gains in the language and music domains [22,36].…”

“…It makes computation infeasible for even moderately large systems. Moreover, complex molecular properties, such as predicting the yield of chemical reactions, are still beyond the reach of what is typically referred to as computational chemistry methods [1]. Instead, these properties have to be extrapolated from an often small experimental dataset [2,3].…”

The transferability limits of static benchmarks

“…However, it is usually not straightforward to quantify a method's uncertainty for a specific case under consideration. [1][2][3][4] Owing to the lack of analytical results for error estimation, the reliability of quantum chemical methods is assessed by numerical benchmarking. The error with respect to some reference data is determined for a predefined set of molecules.…”

The long road to calibrated prediction uncertainty in computational chemistry

“…The main error sources in CC have been reviewed recently 1,2 , and can be tagged as numerical, parametric, and model errors. I come back briefly on these categories in order to discuss the expected errors distributions, which is an essential ingredient to define UQ validation methods.…”

“…As stated in recent perspective articles [1][2][3] , uncertainty quantification (UQ) in computational chemistry (CC) is still in its early stages of development. For instance, in electronic structure theory, at the exception of the BEEF-type methods 4 , none of the methods implemented in popular CC codes provides an uncertainty or a confidence index on the calculated properties.…”