A Unified Active Learning Framework for Designing Energy-Relevant Molecules

Solar photovoltaic (PV) technology has received immense, global interest in recent years. Many novel devices have been created beyond conventional silicon solar cells, such as thin film PV, organic PV, and perovskites. Unfortunately, all such single-junction solar cells suffer from an efficiency cap known as the detailed-balance limit, which limits solar cell efficiency to 33.7%. This limit is primarily due to solar cells unable to absorb light below their bandgap, and inefficiently absorbing light above their bandgap.<br/><br/>One strategy to improve PV efficiency is to use certain organic molecules that up- or down-convert photon energies using interplays between their excited states. Two common types of photon conversion processes are triplet-triplet annihilation (TTA) up-conversion and singlet fission (SF) down-conversion. However, designing efficient TTA and SF molecules comes with several challenges. Namely, the design space of organic photon conversion molecules is massive; for example, 166 billion organic molecules exist with less than 17 atoms. Further, experiments or first-principles simulations to evaluate the excited state energies of these molecules are time- and resource-consuming.<br/><br/>Fortunately, large scale virtual screening and inverse design with machine learning are promising solutions, as they allow accelerated evaluation of properties and efficient exploration of chemical space. Supervised machine learning comes with its own challenges, however. As a primarily data-driven method, it is limited by slow acquisition of labeled data for model training. Further, the chemical design space is diverse, so a large amount of labeled data is required for training to ensure sufficient coverage. Finally, tasks may be very different or require non-overlapping data. For these reasons, a model with high accuracy for desired properties is difficult to obtain.<br/><br/>Therefore, it is useful to develop an efficient, unified strategy for generating training sets and suggesting candidate molecules under different conditions. Active learning is one promising strategy for achieving this, as it efficiently explores chemical space. This study presents an active learning framework for designing energy-relevant molecules, where molecules suitable for TTA up-conversion and SF down-conversion are taken as a case study.<br/><br/>First, an ultra-fast chemical simulation method based on machine-learned calibrations to tight binding is developed for accelerating the labeling process. The calibration training set is carefully curated to ensure both breadth of chemical space and depth in space of molecules of interest, namely large aromatic molecules with pi-conjugated bonds.<br/><br/>Next, we use this accelerated labeling to generate a large molecular database and benchmark various active learning strategies with different priorities over this database. Namely, we implement different acquisition functions for both pool-based and generative-based active learning approaches, based on suitability terms developed for target tasks, uncertainty terms derived from surrogate models, and domain knowledge terms for different specific applications.<br/><br/>Finally, a generative machine learning model is developed based on the informative database derived from the unified active learning framework. The generative model considers both suitability and synthetic accessibility to reduce experimental effort. A pool of molecules suitable for TTA up-conversion and SF down-conversion are proposed, verified with higher-fidelity computational chemistry methods, and finally demonstrated with experiments.<br/><br/>Overall, a unified active learning framework is developed, and molecules suitable for TTA and SF are proposed. The suitability of these materials as photon conversion materials to improve PV efficiency is a promising demonstration of the utility of this approach for designing energy-relevant molecules.