Research | Parikshit Pareek

At P²-LAB we work where power systems, optimization and machine learning meet. Our aim is thoughtful AI: physics-inspired methods that keep grid operation and planning secure as networks turn decentralised and renewable-heavy.

Areas: structured-data FMs AI for power systems learning for optimization power systems & other

Structured-Data Foundation Models

4 papers

Labelled context rows, unlabelled query rows. Each query reads the table (thicker arc, more attention) and its y is filled in one forward pass, with no training. Square size is the value.

Prior-data fitted networks that learn in context. With Decoupled-Value Attention, one pretrained model solves power flow on network configurations it has never seen — no retraining, no gradient updates — and the same idea carries to tabular science beyond the grid.

Supported by ANRF PMECRG 2025–2028.

Selected papers

Decoupled-Value Attention for Prior-Data Fitted Networks: GP Inference for Physical Equations arXiv, 2025
Gradient-Free Topology Adaptation for Power Flow Surrogates via In-Context Whitening arXiv, 2026
Mechanistic Evidence for Spectral Structures in Prior-Data Fitted Networks arXiv, 2026

show all papers in this area ›

AI for Power Systems

18 papers

Uncertain loads and PV pass through a closed-form surrogate. Every node’s voltage envelope appears at once, with no Monte Carlo, and nodes that can breach 0.95 p.u. are flagged.

Gaussian-process surrogates that turn power flow into a closed-form, differentiable map, V = f(s). From it come probabilistic voltage envelopes, contingency screening in seconds and guarantees on voltage risk — without Monte Carlo.

Selected papers

Learning Power Flow with Confidence: A Probabilistic Guarantee Framework for Voltage Risk Electric Power Systems Research, 2027
Gaussian Processes in Power Systems: Techniques, Applications, and Future Works Applied Energy, 2026
A Framework for Analytical Power Flow Solution Using Gaussian Process Learning IEEE Trans. Sustainable Energy, 2022

show all papers in this area ›

Learning for Optimization

6 papers

Same start, same optimum. Blue: a learned proxy, one pass. Black: an iterative solver, step by step.

Learned optimization proxies that return near-optimal, physically feasible dispatch from only a few labelled solves — and language models that write hard MILP instances and tune hyperparameters themselves.

Selected papers

Optimization Proxies using Limited Labeled Data and Training Time — A Semi-Supervised Bayesian Neural Network Approach ICML, 2025
Teaching LLMs to Generate Challenging MILP Instances via Solver Feedback arXiv, 2026
Small LLMs with Expert Blocks Are Good Enough for Hyperparameter Tuning ICLR Workshop, 2026

show all papers in this area ›

Power Systems & Other Works

Optimal power flow and AC feasibility, small-signal stability, TSO–DSO flexibility and storage, and a realist’s look at quantum power flow.

Recent: Limitations of Fault-Tolerant Quantum Linear System Solvers for Quantum Power Flow; Towards AC Feasibility of DCOPF Dispatch; Convexified Flexibility Area Identification at TSO–DSO Interface

show all 25 papers ›