Quantum-enhanced machine learning for predictive catalysis: A unified framework for transition metal complex activity forecasting

The presented paper illustrates a new machine learning-augmented paradigm based on quantum use to predict the activity of transition metal catalysts. The strategy combines hierarchical graph neural networks with parameterized quantum circuits to capture complexities of d-orbital interaction and of electron correlation. The framework also consists of quantum embeddings, attention mechanisms, and multi-tasks on learning objectives to generalize to unseen reaction classes. The learned embeddings provide an unprecedented predictive ability (mean absolute error: MAE = 0.94 kcal/mol), which is more than a 47 percent improvement over state of the art methods, but still have R 2 = 0.96 and Spearman 0.97. On virtual screening process 1.8 million complexes of transition metals provided 71 percent hit compared to conventional methods (10-30 percent). The efficacy of the paradigm of using quantum embeddings in rational catalyst design is justified by the demonstration in ablation studies where quantum embeddings lead to a 24% gain in performance.