Accelerating Chemical Discovery with Machine Learning: Simulated Evolution of Spin Crossover Complexes with an...

Area of Science:

• Computational Materials Science
• Inorganic Chemistry
• Machine Learning Applications

Background:

• Materials discovery traditionally relies on time-consuming simulations.
• Machine learning (ML) offers a faster alternative for evaluating material properties.
• Accurate prediction of energies and properties is crucial for identifying novel materials.

Purpose of the Study:

• To accelerate the discovery of unconventional spin-crossover complexes.
• To integrate genetic algorithm (GA) optimization with artificial neural network (ANN) predictions for efficient materials exploration.
• To develop an error-aware ML-driven strategy for materials discovery.

Main Methods:

• Utilized a genetic algorithm (GA) for optimization.
• Employed an artificial neural network (ANN) for predicting spin-state splitting in inorganic complexes.
• Explored a chemical space of over 5600 candidate materials using eight metal/oxidation state combinations and a 32-ligand pool.
• Implemented an error-aware ML strategy to guide the GA within the ANN's reliable prediction domain.

Main Results:

• Discovered 80% of potential lead materials by limiting GA exploration near ANN training points.
• Achieved average unsigned errors of 4.5 kcal/mol on a 51-complex subset, close to the ANN's baseline error of 3 kcal/mol.
• Significantly reduced lead identification time from days (DFT-driven GA) to seconds (ANN-driven GA).

Conclusions:

• ML, particularly ANNs combined with GAs, dramatically accelerates inorganic material discovery.
• The error-aware ML strategy enhances the efficiency and reliability of discovering novel materials.
• This approach demonstrates a powerful paradigm shift for computational materials science, moving towards faster and more accurate discovery cycles.