From Atoms to Industry: Multiscale-AI Hybrid Approach for Advanced Process Control
Multiscale modeling
Our group pursues a multiscale approach to understand, predict, and control a wide range of complex chemical and energy processes. Chemical processes inherently span multiple time and length scales, from atomic-level interactions to particle dynamics and macroscopic transport phenomena. It should be noted that there is no one-size-fits-all approach, and various methods are integrated to address this. First, at the molecular scale, density functional theory (DFT) can provide atomistic simulations that reveal elementary reaction mechanisms and yield accurate kinetic parameters across complex reaction pathways. Moving beyond the atomistic level, the kinetic Monte Carlo (kMC) algorithm offers a stochastic simulation framework to simulate the system’s temporal evolution, executing the potential events over extended time scales and providing high-fidelity system output to understand the system in detail. Further, the continuum-scale models efficiently represent particle ensembles by grouping them into classes and tracking how their distributions evolve in response to nucleation, growth, aggregation, or breakage events.
Hybrid modeling
The above ones are first-principles (FP) models based on the actual physics, such as mass/energy balances, thermodynamics, kinetics, etc. However, it requires having appropriate system knowledge. On the other hand, machine learning (ML) models act like a black-box model, giving more flexibility and helping identify the model in various fields. To train the model, however, a large amount of data is necessary. The hybrid model (HM) becomes a promising solution by integrating system-independent physics and system-specific process data. Moreover, it is known that the HM strategy can capture latent mechanisms of complex processes.
Process control
The ultimate goal of our modeling strategies is to translate scientific insights into practical tools for process control and optimization. By leveraging hybrid models that combine FP knowledge and data-driven learning, we enable real-time decision-making and robust control of complex processes.
Machine Learning and Applications to Chemical Processes
Our group has developed an innovative AI model called CrystalFormer that can predict crystallization characteristics using molecular structures. By training a transformer-based language model on 1.8 billion molecules, we interpret SMILES strings as a “universal chemical language” and combine Bayesian regression with physics-based population balance equations to predict thermodynamic and kinetic properties such as solubility and growth rates. Validated on paracetamol and salicylic acid, our system achieves over 98% accuracy in solubility predictions and under 8% error in growth parameters, and can forecast crystal size distributions with confidence intervals without experiments, offering a groundbreaking tool for pharmaceutical Go/No-Go decisions and process optimization.