The Photoswitch Dataset: A Molecular Machine Learning Benchmark for the Advancement of Synthetic Chemistry

Abstract

<jats:p>The space of synthesizable molecules is greater than $10^{60}$, meaning only a vanishingly small fraction of these molecules have ever been realized in the lab. In order to prioritize which regions of this space to explore next, synthetic chemists need access to accurate molecular property predictions. While great advances in molecular machine learning have been made, there is a dearth of benchmarks featuring properties that are useful for the synthetic chemist. Focussing directly on the needs of the synthetic chemist, we introduce the Photoswitch Dataset, a new benchmark for molecular machine learning where improvements in model performance can be immediately observed in the throughput of promising molecules synthesized in the lab. Photoswitches are a versatile class of molecule for medical and renewable energy applications where a molecule's efficacy is governed by its electronic transition wavelengths. We demonstrate superior performance in predicting these wavelengths compared to both time-dependent density functional theory (TD-DFT), the incumbent first principles quantum mechanical approach, as well as a panel of human experts. Our baseline models are currently being deployed in the lab as part of the decision process for candidate synthesis. It is our hope that this benchmark can drive real discoveries in photoswitch chemistry and that future benchmarks can be introduced to pivot learning algorithm development to benefit more expansive areas of synthetic chemistry.</jats:p>

Choosing the right molecular machine learning potential

“…This point becomes particularly relevant if one considers that KM-GD models (KRR- In the case of kernel models, specialized software programs have been developed to enable such an efficiency-boosting. 103,104 However, for comparing timings on the ethanol data set, we performed all simulations on CPUs, using the same computer hardware architecture and number of cores to allow for a fair comparison between different models, alleviating possible dependencies of the computational timings concerning specific hardware configuration.…”

Machine Learning for Electronically Excited States of Molecules

“…The problems of TDDFT have been discussed very recently by Thawani et al. 548 The authors developed a data set for relevant photoswitches, which are useful, e.g., for medical applications or renewable energy technologies. To this aim, photochemical properties of azobenzenes and associated derivatives were manually extracted from experimental papers.…”

Molecular excited states through a machine learning lens

“…Moreover, ML can also learn experimental intermediate properties, such as the absorption wavelength of photoswitch molecules. 215 QM methods could also be used to calculate this property (and ML would be then used as a surrogate model for QM), but these methods can be computationally demanding and can lead to less accurate data than those obtained with experimental measurments. 215 Learning intermediate properties may also provide in-depth theoretical insight into the photophysical processes that takes place in the material and that can be experimentally measured.…”

“…215 QM methods could also be used to calculate this property (and ML would be then used as a surrogate model for QM), but these methods can be computationally demanding and can lead to less accurate data than those obtained with experimental measurments. 215 Learning intermediate properties may also provide in-depth theoretical insight into the photophysical processes that takes place in the material and that can be experimentally measured. Therefore, ML can use experimental data as input and predict important intermediate properties, thus directly linking experiments to the theoretical description.…”

“…For example, ML models that use 3D structural descriptors for the calculation of ground-state energies often serve as the basis for models for the calculation of excited-state energies, electronic couplings 172,207,[209][210][211]213 and absorption wavelengths (see the previous sections and Box 3). 215 The main advantage of learning intermediate rather than ultimate properties is that it is usually easier to generate additional reference data for the former, opening much more efficient and economical ways for materials design. One of these ways is to use unsupervised learning to determine the most important structural motives, enabling the design of materials with desired intermediate properties.…”