Multimodal machine learning model increases accuracy

Building on previous breakthroughs that showcased a transformer-based language model’s capacity to predict adsorption energy using only human-readable text, researchers from Carnegie Mellon University’s Department of Mechanical Engineering have pioneered an innovative methodology that incorporates multimodal learning. This approach links various model configurations and significantly boosts the language model’s proficiency in performing prediction tasks without the need for task-specific labels. Their novel process reduces the mean absolute error of energy prediction for adsorption configurations by 7.4-9.8%.