Dr. Harsha Vardhan Tetali
Postdoctoral Researcher | Department of Computer Science | University of Helsinki
Postdoctoral Researcher | Department of Computer Science | University of Helsinki
Welcome to my website! I am Harsha Vardhan Tetali, a Postdoctoral Researcher in the Department of Computer Science at University of Helsinki. My present research focuses on working on machine learning models for metabolic models.
I am passionate about research involving mathematical ideas in machine learning and signal processing. I previously worked on modeling of SSD NAND channel modeling at Marvell Semiconductor, Inc. My PhD work mainly focused on physics-informed matrix factorizations. My work has been published in IEEE Transactions on Signal Processing.
When I'm not in the lab or classroom, I enjoy long walks around the city of Helsinki.
My research integrates mathematical concepts from various fields into machine learning, with a focus on physics-informed approaches and signal processing applications.
Currently developing machine learning models for metabolic systems at University of Helsinki, exploring computational approaches to understand cellular metabolism.
My PhD research focused on incorporating physical constraints and domain knowledge into matrix factorization algorithms for improved performance and interpretability.
Industrial research experience at Marvell Semiconductor involving mathematical modeling of storage device channels for improved performance and reliability.
My research ideology is built on the foundation of mathematically rigorous approaches to machine learning problems. I investigate how domain-specific knowledge can be incorporated into learning algorithms to improve both performance and interpretability. This work contributes to the broader field of physics-informed machine learning and has practical applications in engineering and computational biology.
My research focuses on developing interpretable physics-informed machine learning algorithms for structural health monitoring (SHM). The central aim is to enhance interpretability within physics-informed machine learning (PIML) by leveraging classical machine learning algorithms that are better suited for interpretability in critical applications.
This optimization framework incorporates physical constraints through regularization, where y represents data from a physical process, F is a machine learning model, M is a masking operator, and T represents a partial differential operator with coefficients c.
Developing novel methodologies that extend beyond Physics-informed Matrix Factorization into more sophisticated mathematical domains. This includes exploring constraint-based approaches in differential geometry, particularly involving manifold structures, and investigating how physical constraints can be meaningfully incorporated into abstract mathematical frameworks.
Advancing the development of more transparent and reliable AI systems for physics-based data. Key focus areas include designing architectures that maintain high performance while reducing data requirements through intelligent incorporation of domain knowledge and physical constraints, bridging the gap between traditional physical modeling and modern machine learning approaches.
Conducting fundamental research in deep learning interpretability with emphasis on developing theoretical foundations for neural networks. This includes investigating mathematical principles underlying neural network behavior, analyzing approximation capabilities in various function spaces, and establishing rigorous frameworks for understanding learning dynamics and generalization properties.
Developing matrix factorization models combined with wave operators for structural health monitoring. This approach enables interpretability by recovering distinct wave data modes and successfully identifying relevant artifacts in regions with missing data, particularly valuable for non-destructive evaluation applications.
Exploring how wave-informed matrix factorization exhibits data-driven approximate eigen-decomposition properties. This enables extraction of vectors that closely approximate eigenvectors of wave operators, with approximation errors guided by input data characteristics rather than finding exact eigenvectors.
Applying physics-informed approaches to environmental monitoring and infrastructure assessment. This includes monitoring critical infrastructure from power plants to renewable energy systems, where the framework's ability to capture nonlinear behaviors while maintaining computational efficiency and interpretability is crucial.
My research encompasses three interconnected theoretical frameworks:
Looking ahead, I am particularly interested in extending these theoretical frameworks to more sophisticated mathematical landscapes, especially in nonlinear partial differential equations, manifolds, and related fields. This extension presents opportunities to explore how physical constraints and interpretability principles can enhance our understanding of complex mathematical systems.
My vision involves developing new theoretical tools that bridge the gap between physics-informed approaches and advanced mathematical structures, potentially leading to novel insights in both theoretical and applied domains. To fully realize this vision, I seek collaborations with researchers in functional analysis, differential geometry, and theoretical optimization.
My research contributes directly to educational initiatives, with theoretical frameworks and algorithms now serving as foundational material in graduate-level Physics-informed Machine Learning courses. This demonstrates the pedagogical value and broader educational impact beyond technical contributions.
Practical applications include Department of Energy-funded projects for non-destructive evaluation of nuclear fuel rod cladding, leveraging wave-informed matrix factorizations to create "digital fingerprints" of ceramic composite materials for rapid structural integrity assessment.
My research has been published in top-tier journals and conferences across signal processing, machine learning, and structural health monitoring.
For a complete and up-to-date list of publications, please visit my Google Scholar profile.
Conducting research on machine learning models for metabolic systems in the Department of Computer Science.
Worked on advanced SSD NAND channel modeling, developing mathematical models for storage device optimization.
Contributed to channel modeling research for solid-state drive optimization.
Conducted PhD research on physics-informed matrix factorizations while assisting with undergraduate courses.
Pursued M.Tech degree while assisting with research projects and teaching responsibilities.
Download my complete CV: Curriculum_Vitae_2025_Summer.pdf
Machine Learning, Signal Processing, Physics-Informed Methods, Matrix Factorizations, Computational Biology, Storage Systems
Programming: Python, MATLAB, C++
Machine Learning: Deep Learning, Matrix Factorization, Physics-Informed Neural Networks
Signal Processing: Channel Modeling, Optimization, Statistical Methods
Tools: TensorFlow, PyTorch, NumPy, SciPy
IEEE Member, Professional associations in signal processing and machine learning