MRS Fall Meeting 2020

Automatic Cyclometalated Iridium(III) Complex Design Using a Data-Driven Fragmentation Representation of Molecules

View publication


Data is a resource to drive discovery, fueled by multiple factors such as new funding, open source, and AI. Inspired by demands from the materials industry, an inverse design solver accompanied with an end-to-end AI-based tool for small organic molecule design is developed. Our published studies show good correspondence between predicted properties and calculated values [1]. In this report, we extend the scope of our tool, ensuring that the newly generated ligands form optimized Ir(III) complex structures, and finally perform validation through DFT simulations confirming energy gaps large enough to blue light LED material candidates. Cyclometalated Ir(III) complexes are a well-established class of metal-organic compounds with attractive photophysical properties [2]. The design of the ligands around the Ir ion dominates the performance. By using open data sets and reliable models, a vast number of ligand candidates can be generated. However, considering the complexity of the three-fold symmetric structure, it is still difficult to effectively calculate representative characteristics even when subject matter expert insight is included. An automatic, high throughput workflow is discussed in the followings. Initial searches extracted 1400 samples with CIE 1931 color coordinates and SMILES information. For deep blue emission the target values are (0.15, 0.12). This corresponds to a calculated energy gap larger than 2.5 eV according to experience. Multiple fused rings in the ligands leads to encode features with atom count, ring count, substructures, and fingerprints. In the prediction step, the hyper-parameters are further optimized by exhaustive grid search algorithm. The model with highest cross validation score is input to estimate new features and ligand generation. In parallel, another framework is established to inherit the generated results. Newly generated ligands are converted to 3D structures and optimized by force field. Although the tris-heteroleptic Ir(III) complex is the more general use case, the detail of dealing with tris-homoleptic ones is described here. The two coordinating ions (A and B) to Ir ion are identified by a rulebased algorithm with more than 90 % accuracy. The three ligands coordinates are put in a 3D space without interference and Ir ion is located at the origin. The atom and bonding information are updated and do 2nd structure optimization. Because the ligand structures are optimized prior to the attachment, the degrees of freedom can be reduced to only three for the purpose of finding energy in global minimum. Three ligands are rotated simultaneously along the vector sum of Ir-A and Ir-B. By using genetic algorithm with many enough iterations, the angle combination with the lowest total energy is obtained. Finally, the hydrogen atoms are attached and ready for DFT simulation. The workflow is validated by reproducing the numbers in [3]. There are 100 samples with tabulated values of LUMO, HOMO, dipole moment, S1, and T1 wavelength. Different software and parameters (G98 vs. GAMESS) are used. The results show good correspondence with R-square scores: 0.98, 0.98, 0.74, 0.71, and 0.89. With this framework, we can validate the newly generated ligands also have large enough band gap energy such as c1cnc2ccn3cc(C4CCCCC4)nc3c2c1 (4.20 eV), c1ccc2c(c1)ncn1c3ccccc3nc21 (3.64 eV). The framework is capable of forming facial and meridional coordination geometries by adjusting the fitness in the optimization algorithm. Multiple dockerized packages provide such workflow automation. The potential of the developed tool is still increasing by adapting the customization for other functional materials. [1] S. Takeda, et al. “Molecular Inverse-Design Platform for Material Industries”, Proc. of KDD (2020) [2] Z. Wu, et al, “Novel Design of Iridium Phosphors with Pyridinylphosphinate Ligands for High-Efficiency Blue Organic Light-emitting Diodes”, Scientific Reports (2016) [3] D. B. Knowles, U.S. Patent 2008/0297033A1