Research & Projects

The key focus of my Ph.D. thesis is on developing novel AI/ML and optimization techniques for process systems. Specifically, I work on:

Novel algorithms for data-driven optimization for high fidelity simulations or black-box process systems
Multi-fidelity hybrid models that integrate machine learning surrogate models with first-principle models/domain specific knowledge.
Hybrid modeling strategies that integrate sparse and symbolic regression to learn system dynamics while enhancing interpretability with explainable AI tools.

A data-driven spatial branch-and-bound algorithm for optimizing black-box functions - DDSBB

A current open challenge in the mathematical programming field is optimization with black-box models and/or data-driven models. These models pose a fundamental difficulty as they lack closed-form analytical expressions, rendering conventional optimization algorithms ineffective or even inapplicable. To address this challenge, I have co-developed an optimization algorithm which uses advanced mathematical programming concepts such as underestimating functions, spatial branch-and-bound heuristics, lipschitz continuity etc., and state-of-the-art ML methods.

I have also conducted a rigorous convergence analysis for a special class of Lipschitz-continuous systems, strengthening the theoretical foundation of the algorithm.

The initial beta version, V1.0, has been released as open-source software on GitHub, and can be found here. The updated version V2.0 is already in the pipeline, will be released soon!

Relevant publications

Suryateja Ravutla, and Fani Boukouvala, "Data-Driven Lipschitz-Informed Convex Underestimators for Branch-And-Bound Optimization of Black-Box Functions." Journal of Global Optimization, 2025, under review.

Suryateja Ravutla, Jianyuan Zhai, and Fani Boukouvala, "Hybrid Modeling and Multi-Fidelity Approaches for Data-Driven Branch-and-Bound Optimization." In Computer Aided Chemical Engineering, vol. 52, pp. 1313-1318. Elsevier, 2023.

Try it out here! This is a small demo on basic implementation on PyDDSBB. More details in the blog!

Hybrid Multi-fidelity Surrogate Modeling approaches for Adaptive Sampling and A-priori Sampling Optimization Techniques

While there is a rapid deployment of AI/ML systems in all fields, bias may arise as an unintended consequence, due to data-dependent parameters of the models. Many AI and ML models, especially deep learning models, are often considered black boxes. Developing a good model that is accurate and generalizable is a challenge when the sampling is limited and the sampling costs are high. This is often the case in most high-fidelity simulations!

Hybrid modeling and physics-informed machine learning can play a crucial role in addressing these challenges. I developed multi-fidelity hybrid modeling frameworks that combine machine learning surrogates with simulation data and domain knowledge to enable accurate modeling and robust optimization of complex process systems. Specifically, I demonstrated that low-fidelity system data can be systematically leveraged through model-corrective hybrid surrogates to improve prediction accuracy and optimization performance.

Relevant publications

Suryateja Ravutla, and Fani Boukouvala, "Effects of Surrogate Hybridization and Adaptive Sampling for Simulation-Based Optimization." Industrial & Engineering Chemistry Research, 2025 64(18), pp.9228-9251.

Effects of Surrogate Hybridization and Adaptive Sampling for Simulation-Based Optimization

I have also investigated the effect of surrogate hybridization on optimization performance and computation time. Specifically, I explored two common approaches for optimizing simulation/black-box functions:

Surrogate-based optimization – samples are collected a priori, an ML surrogate model is built, and classical solvers are used to optimize it
Adaptive sampling-based optimization – the search space is explored adaptively, with samples collected as needed to identify the optimum.

In this work I have identified conditions under which surrogate-based methods (with 3 different type of formulations – reduced space, full-space and ReLU formulations) or adaptive sampling methods are more effective, guiding the formulation choice, for robust global optimization

Relevant publications

From Hybrid Models to Interpretable White-Box Models: Data-Driven Process Model Discovery

In my most recent work, I extended hybrid modeling techniques to enable data-driven model discovery. The goal is work is to uncover unknown/missing physics and build interpretable models directly from data for dynamic process models. I developed two novel strategies that combine hybrid modeling with sparse regression, enabling simultaneous learning of system dynamics and governing equations. To enhance interpretability, I integrated explainable AI (XAI) tools and symbolic regression into the hybrid modeling framework to help analyze, refine, and improve the black-box components of the model.

Relevant publications

Suryateja Ravutla, and Fani Boukouvala, "A Two-stage Hybrid Modeling Approach for Data-Driven Process Model Discovery." In preparation.

Suryateja Ravutla, and Fani Boukouvala, "Integrating Hybrid Modeling and Multifidelity Approaches for Data-Driven Process Model Discovery." LAPSE, Systems and Control Transactions (2024), Volume 3: pg 351 - 358

A Machine Learning Approach to Material Screening, Process Modeling and Optimization

During my time at Georgia Tech, I engaged in a range of projects collaborating with peers from diverse backgrounds and researchers from various universities. One of the projects was focused on the simultaneous optimization of process conditions and adsorbent selection for a modular Vacuum Pressure Swing Adsorption (VPSA) system targeting CO₂ capture. We applied surrogate-based NLP and MINLP optimization methods to evaluate 75 candidate adsorbents. I developed an ML framework to classify adsorbent feasibility using a support vector classifier with dimensionality reduction techniques such as PCA, and to predict corresponding process performance using a multivariate neural network.

In another project, I contributed to a multi-objective optimization study of a novel adsorbent, Elastic Layer-structured Metal-organic Framework-11 (ELM-11), for CO₂ capture via a VPSA process. I led the design of experiments (DoE) sampling strategy.

Relevant publications

Yuya Takakura, Suryateja Ravutla, Kim Jinsu, Keisuke Ikeda, Hiroshi Kajiro, Tomoyuki Yajima, Junpei Fujiki, Fani Boukouvala, Matthew Realff, Yoshiaki Kawajiri, "Surrogate model optimization of vacuum pressure swing adsorption using a flexible metal organic framework with hysteretic sigmoidal isotherms." International Journal of Greenhouse Gas Control 2024, 138, p.104260.

Kim, Sun Hye, Héctor Octavio Rubiera Landa, Suryateja Ravutla, Matthew J. Realff, and Fani Boukouvala, "Data-driven simultaneous process optimization and adsorbent selection for vacuum pressure swing adsorption." Chemical Engineering Research and Design 188 (2022): 1013-1028

Process Modeling and Parameter Estimation: Leveraging Derivative-Based and Stochastic Optimization Algorithms

As an undergraduate researcher at the Indian Institute of Technology Hyderabad, I worked on computationally developing a physics-based model for the calcium dynamics in hippocampal neurons, by understanding the calcium signaling pathway. This model served as a basis for understanding how neurons communicate using calcium signals, providing valuable insights into their functioning. A few other projects I worked on involved modeling biological system pathways and modeling of cell-to-cell variability in viral infection.

Relevant publications

Saxena, Abha, Suryateja Ravutla, Kishalay Mitra, Jayanta Chakraborty, David Murhammer, and Lopamudra Giri,"Evolution of a single-cell predictive model for packaging and budding of viruses based on TEM based measurements." Authorea Preprints (2021).

Saxena, Abha, Suryateja Ravutla, Vikas Upadhyay, Soumya Jana, David Murhammer, and Lopamudra Giri, "Statistical modeling of cell-to-cell variability in viral infection during passaging in suspension cell culture: Application in Monte-Carlo simulation." Biotechnology and bioengineering 117, no. 5 (2020): 1483-1501.

Vikas Upadhyay, Suryateja Ravutla, Vaibhav Dhyani, Kevin George, Sarpras Swain, Kishalay Mitra, and Lopamudra Giri, "A model screening framework for the generation of Ca2+ oscillations in hippocampal neurons using differential evolution." The 9th International IEEE/EMBS Conference on Neural Engineering (NER), pp. 961-964. IEEE, 2019.

Sarpras Swain, Sathish Ande, Suryateja Ravutla, Soumya Jana, and Lopamudra Giri, "Spatially resolved calcium spiking in hippocampal neurons: Estimation via confocal imaging and model-based simulation." The 8th International IEEE/EMBS Conference on Neural Engineering (NER), pp. 279-283. IEEE, 2017

Page updated

Google Sites

Report abuse