Research
Intelligent Multiscale Process Engineering
Combines microscopic simulation, multiscale models and AI to study material–interface mechanisms, model trustworthiness and detailed fluidization behavior for process optimization, monitoring and intensification.
Research overview
The Intelligent Multiscale Process Engineering research area is jointly led by Dr. Donovan Chaffart and Dr. Yue Yuan and focuses on the intersection of artificial intelligence and chemical engineering. We study how microscopic changes in materials and interfaces affect the macroscopic behavior of reactors and manufacturing processes, and how experimental data and engineering requirements can, in turn, guide material design and process optimization. We combine first-principles calculations, molecular simulation, multiscale models and machine learning to develop hybrid modeling methods that exchange information across scales, balancing prediction efficiency, physicochemical meaning and applicability under practical operating conditions.
Research centers on three themes: intelligent multiscale modeling with information exchange, trustworthy AI, and detailed identification of fluidization behavior. The first connects microscopic mechanisms with macroscopic processes; trustworthy AI examines how models generate predictions and the basis for those predictions; and fluidization research converts complex experimental signals into structural information for process monitoring. Engineering data preprocessing and knowledge constraints run through all three themes. The methods address thin-film deposition, functional interfaces, energy materials and particle fluidization systems to support mechanistic analysis, process optimization and intensification.
Research topics
1. Intelligent Multiscale Modeling with Information Exchange and Material–Interface Processes
Materials and manufacturing processes often span widely different time and length scales. Microscopic simulations can describe surface reactions, molecular migration and interface evolution, but cannot readily cover an entire engineering process. Macroscopic models are computationally efficient, yet may lose microscopic information that determines material properties. We connect these descriptions through intelligent multiscale modeling, allowing microscopic mechanisms and macroscopic operating conditions to inform each other, while using data-driven algorithms to improve computational efficiency.
Methods include density functional theory, molecular simulation, kinetic Monte Carlo, physics-informed neural networks and hybrid models. Yue Yuan focuses on modeling materials and reaction mechanisms under non-ideal conditions. Donovan focuses on moving-interface algorithms, functional surfaces, and the optimization and control of multiscale models. Research systems include thin-film growth, droplet spreading on surfaces and oxygen carriers for chemical looping combustion, with further work directed toward interface problems in energy materials and advanced manufacturing.
In published research on stochastic thin-film deposition, a multiscale physics-informed neural network incorporates gas-phase mass transfer and microscopic surface evolution within a single model. In the validation cases reported in the paper, the trained model achieved computational speedups of approximately 2 and 4 orders of magnitude relative to the SPDE and kMC multiscale models, respectively, providing a more efficient solution method for process parameter optimization.
2. Trustworthy AI and Engineering Data Modeling
Chemical processes place high demands on safety, prediction reliability and mechanistic understanding. We study how AI can provide predictions while explaining their basis, and examine whether the relationships learned by a model agree with engineering knowledge. The team develops a framework for analyzing the transparency of chemical-engineering AI applications along three dimensions: interpretability, informativeness and causality.
Interpretability concerns internal computational logic and information transfer within a model. Informativeness concerns the integration of physicochemical mechanisms, domain knowledge and data. Causality concerns the relationships between inputs and outputs and the basis for their effects. Research includes extracting mathematical expressions from polynomial neural networks, constructing models with knowledge constraints, preprocessing engineering data and integrating data from multiple sources. Particular attention is given to practical problems such as limited data, misaligned measurements and inconsistent data quality.
Published research uses polynomial neural networks to extract continuous, closed-form algebraic expressions, analyze the contributions of network layers and features to the output, and examine the method’s applicability in a fluidized-bed reactor case study. We aim to develop models that support engineering predictions, help identify model bias and understand process behavior, and provide a basis for model improvement.
3. Detailed Identification of Fluidization Behavior and Intelligent Monitoring and Control
Gas–solid fluidized beds are widely used for reactions, separation and particle processing. Bubbles, particle clusters and local concentration changes jointly affect transport efficiency and reaction performance. Conventional averaged parameters cannot fully describe these rapidly changing heterogeneous structures. We combine pressure and optical-fiber measurement signals with numerical models to study instantaneous flow behavior and flow-regime transitions, providing a more detailed description for process intensification.
Research focuses on extracting features from fluctuating signals, labeling flow behavior through unsupervised learning, coupling multiscale flow models with physics-informed neural networks, and predicting flow and reaction performance during fluidized-bed scale-up. Yue Yuan focuses on identifying detailed fluidization behavior from experimental signals, while Donovan focuses on intelligent algorithms and hybrid model development. Together, they advance cross-validation among data, theory and experiments.
Under the research conditions established to date, we have reconstructed instantaneous internal states from low-dimensional signals and are conducting fine control tests based on the reconstructed states. Further work combines multiscale mechanistic models with active learning to study constrained adjustments to gas flow, temperature, feed and other operating conditions. Models and parameter recommendations are updated after changes in feedstock, product or operating conditions, with the aim of connecting state sensing, process optimization and execution feedback.
Applications address high-value fine particulate materials, including electronic-grade spherical SiO₂, SiC fine-powder processing and particle coating. The focus is on relating reactor flow regimes, local heterogeneous structures and particle histories to particle-size distributions, surface states and quality consistency. Through adaptation to target materials, reactor scale-up and integrated online testing, we explore process packages that support stable continuous production. Specific performance and applicability require joint validation with the target materials and equipment.