Accelerating Photofunctional Molecule Discovery with Artificial Intelligence.

With rapid advances in computer hardware and algorithms, artificial-intelligence-assisted (AI-assisted) computational chemistry provides a new pathway for the design of new chemical compounds. In this issue of ACS Central Science, Masato Sumita, Koji Tsuda, et al. utilize neural-network-based AI and quantum mechanical computations to develop ready-to-synthesize organic photofunctional molecules with desired excitation energy.[1]

Discovery of new organic materials such as molecular systems and polymers, in general, involves several steps, including candidate search in vast and complex chemical spaces, optimization of synthetic routes, and characterization of the samples (Figure ). While these nuts and bolts of molecular design often make this process nontrivial, none of these elements can be omitted. Although often restricted or limited, the acquisition of a comprehensive data set that covers a wide area of the relevant or acceptable material space not only enables direct material selection prior to the experiment but also greatly facilitates data exploration such as trend mining, classification, and correlation analysis.[2] Several approaches such as trial-and-error-based search, random selection, intuition-driven design, high-throughput computations or experiments, brute-force enumeration, and genetic algorithms have been attempted to find candidate materials with desired properties in a given chemical space. Recent advances in AI technologies including conventional machine learning and deep learning have led to a significant leap in accelerated predictions of properties and on-demand design of materials.[3] The usefulness of AI-assisted approaches is increasingly being accepted by researchers as demonstrated by several recent successful examples.[4−6]

Figure 1

Artificial intelligence driven by data-centric informatics accelerates the design of new materials and prediction of their properties.

Sumita and co-workers aimed to construct a material design platform using advanced AI techniques and to apply it to designing new organic photofunctional molecules. Photofunctional molecules are well-known as smart materials that respond to external light stimuli. After the introduction of materials with photofunctional properties, interest in such materials has rapidly increased in the fields of bioimaging, chemical sensing, electrochromic devices, optical filters, organic electronics, and photofunctional sensors.[7,8] Naturally, the demand for newer synthesizable organic materials with better stability, higher reliability, and proper excitation energies for specific applications is steadily growing. The authors have made the search for new molecules more efficient using AI and computational chemistry.

This work has four key stages: (1) AI model training for the virtual synthesis of organic molecules, (2) generation of organic molecules using the AI model, (3) screening for a target property (excitation wavelength), and (4) validation by experimental characterization. Trained on 130 000 benchmark molecules made of C, H, O, and N moieties, the AI-assisted molecular generator named ChemTS was developed using Monte Carlo tree search techniques and recurrent neural networks.[7] ChemTS was coupled with the density functional theory (DFT) simulation package, Gaussian 16,[9] to generate 3200 unique photofunctional molecules with simulated property values in just 10 days. The authors have identified 86 potential photofunctional molecules with a targeted excitation wavelength (200, 300, 400, 500, and 600 nm, respectively). The most significant result of this work is the successful synthesis of six new molecules from 86 candidates. The absorption spectra obtained by experimental ultraviolet visible spectroscopy and time-dependent DFT revealed five finalists with target excitation wavelengths. Two of these molecules, a derivative of oxazole and a derivative of quinolone, had excitation wavelengths of 207.8 and 301.57 nm, respectively. These photofunctional molecules could be used in applications that require 200 and 300 nm light sources.

The impetus for the data-centric and AI-assisted studies is already becoming popular in several subfields of materials research. Such approaches used for surrogate modeling, analysis of relationships between properties, and molecular classification will become more powerful (in accuracy and robustness) with continuous accumulation of material information and development of computer hardware and algorithms in the future. This study succeeded in revealing five synthesizable photofunctional molecules through the appropriate combination of the AI techniques, quantum mechanical computations, and experiments. Although the focus of this study is on organic molecules for photosensitive applications, the proposed framework can be applied to design of various material systems with target properties.