Speaker
Description
The versatility of organic molecules generates a rich design space for functional materials such as organic semiconductors (OSCs), dyes or molecular switches. Offering unparalleled promise for materials discovery, the vastness of this design space also dictates efficient search strategies, however. This concerns the ability of a search algorithm to explore a large variety of possible solutions while still producing a large ‘hit-rate’ of suitable molecules (i.e. the exploration vs. exploitation trade-off). Furthermore, complex target properties such as internal reorganization energies and optical spectral overlaps are computationally expensive to evaluate, providing an additional incentive towards efficient strategies.
In this context, an active machine learning (AML) approach proved capable of exploring an unlimited search space of potential OSCs through the consecutive application of molecular morphing operations and by iteratively querying predictive-quality first-principles calculations to build a self-improving surrogate model [1]. This optimized AML approach rapidly identified well-known and hitherto unknown molecular OSC candidates with superior charge conduction properties.
In a related project, we critically assessed and systematically improved supervised ML models for reorganization energies of flexible molecules [2]. Here, we found that the use of ML surrogate models for high-throughput virtual screening is not always beneficial, since the models can display poor performance for the desired exceptional candidates. This supports the notion that active or generative models for data selection are preferable to ‘brute-force’ screening.
More recently, we have therefore explored the use of generative ML models for the design of molecular switches with tunable optical addressability. Here, the challenges encountered for OSCs are further aggravated by the difficulty of predicting and optimizing spectra, compared to scalar quantities like the reorganization energy. Furthermore, even less data is available for molecular switches than for OSCs. This latter point is addressed by a transfer learning strategy, which refines a general chemical design model based on iteratively generated data on promising molecular switches.
[1] C. Kunkel, J.T. Margraf, K. Chen, H. Oberhofer, and K. Reuter, Nat. Commun. 12, 2422 (2021).
[2] K. Chen, C. Kunkel, K. Reuter, and J.T. Margraf, Digital Discovery 1, 147 (2022).
| Abstract Number (department-wise) | TH 04 |
|---|---|
| Department | TH (Reuter) |
