Deng Energy and Nanotechnology Group

CRNNs are specialized neural networks that directly encode the Arrhenius and mass action laws. When trained, they not only accurately reconstruct the training data like a standard neural network, but also provide the exact Arrhenius and mass action parameters that describe the reconstructed system. They enable modeling in combustion systems, battery thermal runaway simulations, biological systems, and more, with optional uncertainty quantification capabilities.

For complex energy conversion, chemical engineering and other industrial systems, digital twins can provide detailed information on the distribution of internal complex physical fields and predict their evolution, supporting intelligent operation, diagnosis, and design of the system. By integrating advanced physics-informed neural networks, neural operators, and generative learning methods, our team develops surrogate models for sparse reconstruction, multi-field inference, and dynamic prediction, establishing key foundational models toward multimodal, multifunctional digital twin systems.

KAN-ODEs combine Neural ODEs with Kolmogorov Arnold Networks (KANs). KANs are a new network structure proposed as a highly expressive and interpretable replacement for standard multi-layer perceptrons (MLPs). Coupling them to ODE solvers enables direct learning of dynamical system gradients and hidden physics inference, which we have demonstrated in examples including biological systems, wave and shock formation, complex-valued equations, phase separating systems, and detailed chemistry for combustion systems.

We develop inverse modeling frameworks to extract information about reacting flows from spectroscopy measurements using computer vision principles. We combine physics-based approaches, which ensure accurate treatment of light-medium interactions, with machine learning techniques that reduce computational cost and enhance diagnostic quality. These methods span multiple spatial dimensions, ranging from 0-D chemical reactors to 3-D turbulent flames, enabling the refinement of data variables such as pollutant and temperature, applications include clean energy reacting flows such as ammonia combustion and carbon nanotube synthesis, where accurate, efficient inference is critical for advancing both fundamental science and technology deployment.

The cathode material itself is the primary contributor of battery materials. A long-time and energy-intensive process usually features cathode material production. As a result, the manufacturing cost is a significant contributor to the high cost of cathode materials. To simplify the manufacturing process, we are utilizing a Flame-Assisted Spray Pyrolysis (FASP) method to reduce the synthesis time significantly. With the developed method, we can reduce the production time by more than one order of magnitude.

Replacing the current polycrystalline cathode material with a single-crystal structure is a promising route that slows battery performance fading and improves thermal stability. However, the current methods for single-crystal cathode material manufacturing have a long synthesis time and complex steps, causing difficulties in reducing cost without sacrificing performance. Therefore, we propose a low-cost flame spray synthesis (FSP) method that is simple, fast, continuous, and has scalable potential. Our current project aims to develop a lab-scale FSP method for producing single-crystal cathode materials, such as lithium-nickel-cobalt-manganese oxide and lithium-nickel-cobalt-aluminum oxide. We believe that the proposed method is promising in facilitating the development of cost-effective LIBs with outstanding safety and stability features for energy storage applications.

Energetic materials produce a large amount of heat and gas rapidly, leading to the use of materials for extreme energy such as propellants, explosives, and high-power actuation. Among various types of energetic materials, metal-based energetic nanomaterials offer significant advantages of high energy densities, high reactivity, and tunable properties. Yet, understanding the combustion physics remains poor due to complexity of reaction-transport coupling and nanoparticle dynamics, limiting our ability to control combustion dynamics as desired. In the group, we aim at comprehending these intricate multiphase, multiscale combustion physics using in-situ flame measurements with an ultimate goal of establishing predictive models. Our current interest includes transport-driven flame dynamics control and nano-additive effects on combustion performance.

Characterizing materials’ properties is one of the most important tasks for researchers as it can help understanding of physical behaviors and engineering of materials for specific use cases. However, chemical kinetics and material properties at a high temperature are lacking due to complex multiphase, multiscale physics at a wide range of scales. This challenge makes first-principle-based calculation intractable. We tackle these challenges via an inverse modeling approach in which materials’ properties and chemical kinetics can be directly inferred from in-situ flame measurements. In doing so, we enable quantitative analyses of materials and flame dynamics for various applications including energetic materials combustion, metal fuel reactions, and thermal safety in energy storages.