Research Spotlight:
Professor Somdatta Goswami
Somdatta Goswami is an Assistant Professor in the Department of Civil and Systems Engineering.
Her research focuses on advancing Computational Mechanics, Scientific Computing, and Machine learning, with an emphasis on developing algorithms for AI-accelerated numerical simulations in engineering and biological systems.
Leading the Centrum IntelliPhysics group, she addresses long-time horizon problems and multiscale multiphysics material modeling.
Q: What is the focus of your research?
SG: “My research focuses on advancing Computational Mechanics and Scientific Computing with the help of state-of-the-art machine-learning approaches. To that end, my group works on developing algorithms and architectures within the broad domain of physics-informed machine learning to tackle high-dimensional problems in engineering, physical, and life sciences. We aim to enhance computational efficiency and accuracy in solving complex scientific challenges.”
Q: What initially sparked your interest in Scientific Computing and Computational Mechanics?
SG: “My interest was ignited during my Ph.D. in structural engineering, where I tackled computationally intensive brittle fracture problems. The challenge of accelerating a standard 3D problem from 7 days to 52 hours using physics-informed machine learning models was particularly inspiring. This experience revealed the potential of integrating machine learning with computational methods to solve complex engineering problems more efficiently. Developing a neural operator-based surrogate for generalized crack solutions further demonstrated the power of this approach, solidifying my commitment to pursuing this field in my research career”
Q: What are the key computational challenges in engineering, physical, and biological systems that your research group is addressing? What algorithms and architectures does your group develop to tackle these challenges?
SG: “My research group primarily focuses on efficiently coupling multiple scales in modeling various complex physical processes like failure mechanics and material modeling. We tackle the complex interdependence across time and length scales, where atomic interactions can significantly impact larger-scale phenomena. A major challenge is reducing the high computational costs associated with traditional multiscale modeling techniques. To address these challenges, we develop physics-informed surrogate models to replace the most computationally expensive solvers in multi-scale modeling. Recognizing that surrogate models alone can be inaccurate for dynamic processes, we’ve created hybrid frameworks that couple ML models with numerical solvers, maintaining high accuracy during real-time inference. My group has also developed efficient training strategies for neural operators to operate in high-dimensional spaces, such as physics-informed latent neural operators for modeling fracture. Currently, we’re working on coupling neural operators with numerical methods to provide error-bounded responses in real-time for complex processes, further advancing our ability to model and predict complex physical and biological systems accurately and efficiently.”
Q: How do you utilize AI techniques in numerical simulations to enhance their efficiency and accuracy?
SG “My group employs two primary AI-driven approaches to enhance the efficiency and accuracy of numerical simulations:
a) Physics-aware methods: These methods target specific physical tasks by replacing computational bottlenecks in standard solvers with ML techniques. We maintain information about governing equations while eliminating time-consuming solvers, often incorporating equation-based penalization terms into the loss function. This approach significantly accelerates computations, though it may slightly compromise accuracy for speed. Once trained, these ML models can generalize solutions across multiple parametric conditions, offering versatility and efficiency.
b) Physics-agnostic methods (data-driven approaches): These methods learn feature representations of physical phenomena directly from solutions generated by numerical solvers, without explicit knowledge of underlying equations. This approach is particularly powerful due to its versatility across different physical systems. By learning data-driven representations from previous computations, these methods can rapidly produce outputs for new geometries and conditions. The key advantage is the combination of speed and predictive accuracy, though a potential drawback is that predicted solutions may not strictly adhere to physical laws.
Both approaches aim to strike a balance between computational efficiency and solution accuracy, with each offering unique strengths for different types of problems. Our research focuses on optimizing these methods and developing hybrid approaches that leverage the strengths of both to further improve numerical simulations in various scientific and engineering domains. “
Q: How has computing changed Scientific Computing and Computational Mechanics? How do you see AI and machine learning shaping the future of these fields?
SG: “Computing advancements have revolutionized Scientific Computing and Computational Mechanics, dramatically expanding our problem-solving capabilities. Increased processing power, parallel computing, and high-performance computing infrastructures have enabled researchers to tackle previously intractable challenges. These developments have facilitated more sophisticated numerical methods, enhanced visualization techniques, and specialized software packages, making complex simulations more accessible and insightful across various scientific domains.
Looking ahead, AI and machine learning are poised to further transform these fields. We anticipate AI-assisted problem formulation and optimization and rapid design space exploration through surrogate modeling. Hybrid models combining traditional physics-based approaches with data-driven insights will likely improve accuracy and efficiency. AI is expected to automate mesh generation, accelerate real-time simulations, and facilitate multi-scale modeling across various length and time scales. Physics-informed machine learning models are expected to enhance generalization and interpretability of these models. These advancements will accelerate scientific discovery, enable more accurate predictions in complex systems, and open new frontiers in fields such as materials science, climate modeling, and personalized medicine, pushing the boundaries of computational feasibility in Scientific Computing and Computational Mechanics.”
Q: What are the next steps or future directions for your research?
SG: “My research is expanding in two key directions:
- Hybrid ML-Numerical Solvers: We’re developing solvers that combine pre-trained ML models with numerical methods like FEniCs, Ansys, and Abaqus. This approach aims to drastically reduce simulation times while maintaining reliability, potentially accelerating multiscale simulations from days to minutes.
- ML-Integrated Exascale Simulations: We’re exploring real-time, in-situ training for ML models in high-performance computing environments. By adapting federated learning techniques for exascale simulations, we aim to overcome memory and bandwidth constraints in massive, distributed environments.
These directions seek to enhance computational efficiency and accuracy across scientific and engineering domains, potentially revolutionizing our approach to complex, large-scale simulations.”
Q: What is the impact or value of Rockfish in your research?
SG: “Rockfish has been a crucial asset to my research, particularly in supporting the development of ML models, which require significant GPU resources for training and labeled data generation which runs on multiple CPU cores. The high-performance computing industry is continuously advancing to meet the growing demands of AI-driven workloads, and Rockfish exemplifies this integration by providing state-of-the-art GPUs to cater to the requirements of JHU students and faculty. Furthermore, for a new faculty member, access to such state-of-the-art resources at no cost has been instrumental in accelerating my research.”
One Response
Comments are closed.