Scientific Machine Learning
Learning dynamics from data with structure baked in, neural ODEs/SDEs, physics-informed learning, differentiable simulation, and operator learning / neural operators.
Neural ODEsPINNsNeural OperatorsDifferentiable Sim
Research
Understanding, not just prediction.
Real-world systems, biological, physical, clinical, are dynamical, noisy, and only partially observed. I develop machine-learning methods that respect that structure: models grounded in differential equations and operators, made honest by Bayesian uncertainty, and made explanatory by causality. The goal is not just prediction, but understanding we can trust and act on.
Current focus · PhD thesis
Probabilistic Causal Inference in Complex Dynamical Systems. In the MP-AIX programme, advised by Prof. Peter Benner (Max Planck Institute) and Prof. Philipp Hennig (University of Tübingen).
Themes
Research directions
Probabilistic ML & Bayesian Inference
Treating computation and modelling as inference: Bayesian inverse problems, probabilistic numerics, and calibrated uncertainty quantification.
Bayesian InferenceProbabilistic NumericsUncertainty Quantification
Causal Discovery & Inference
Recovering mechanisms, not just correlations, causal structure learning and treatment-effect estimation in complex, dynamical, confounded settings (my PhD focus).
Causal DiscoveryTreatment EffectsInterventions
Graph Neural Networks & Representation Learning
Structured representations for relational and spatiotemporal systems using GNNs and attention, from multi-agent dynamics to clinical time-series.
GNNsGraph AttentionRepresentation Learning
Machine Learning for Science & Health
Putting the methods to work: electronic health records, medical imaging, and scientific applications where reliability and interpretability matter.
EHRMedical ImagingHealthcare AI
Selected work
GitHub
Projects & code
Synthetic Instrumental Variables
Learning to synthesize valid instrumental variables from observed covariates with a graph-attention network, for less-biased treatment-effect estimation under unobserved confounding (2SLS, evaluated on LaLonde).
Causal InferenceGraph AttentionPyTorch Geometric2SLS
Causal Discovery in Spatio-Temporal Systems
Self-supervised spatio-temporal GNN embeddings feeding a constraint-based (PC) discovery step to recover directed spatial–temporal dependencies beyond Granger causality, on the METR-LA sensor network.
Causal DiscoveryST-GNNPyTorchPC Algorithm
Graph-Informed Signature Kernels
Signature-kernel conditional-independence testing (rough-path theory) combined with GNN-derived biomedical knowledge graphs for sample-efficient causal discovery in clinical time-series (MIMIC-III).
Kernel MethodsRough PathsHealthcare AI
Biologically-Constrained Neural ODEs for Gene Regulation
Continuous-time neural ODEs with biologically-informed sparsity priors on the learned vector field’s Jacobian, recovering interpretable gene-regulatory structure from longitudinal trajectories.
Neural ODEsSystems BiologyCausal Discovery
Decentralized Multi-Agent Path Planning with GNNs
A decentralized navigation policy built on graph attention networks that scales to 100+ agents in simulation and improves throughput over ORCA in dense scenarios.
GNNMulti-AgentReinforcement Learning
Cyclic Stress Ratio Prediction (ML / DL)
ML and deep-learning models predicting the cyclic stress ratio of soils, the basis of a peer-reviewed publication in Environmental Challenges.
Applied MLPublishedScikit-Learn
PET/CT Pulmonary Nodule Detection
A detection pipeline for pulmonary nodules in PET/CT medical imaging, part of ongoing work on reliable ML for diagnostics.
Medical ImagingDetectionPyTorch
Get in touch
Let’s talk research, collaborations, or positions.
I’m always glad to hear from fellow researchers and students, about scientific & probabilistic ML, causal inference, ML for science and health, or potential collaborations.
pawar@mpi-magdeburg.mpg.de