Machine Learning for Science (ML4SCI)

Mentors: Student: Aatmaj28

DeepLenseSim (built on lenstronomy) requires substantial manual intervention for configuring parameters, managing outputs, and iterating on failures across multi-step simulation pipelines, creating bottlenecks in large-scale dataset generation and limiting parameter space exploration. I will build an Agentic AI framework following the HEPTAPOD philosophy to autonomously orchestrate DeepLenseSim workflows across three phases: (1) implementing schema-validated Pydantic tools wrapping the DeepLens class for all four model configurations, an LLM orchestration layer via OpenRouter function calling, human-in-the-loop confirmation flow, and a compatibility layer addressing pyHalo/lenstronomy/numpy version conflicts; (2) building a ParameterScanTool for automated multi-point sweeps (axion mass, halo mass, redshift), a ValidationTool for automated image quality checks, and exploration of RL-based parameter optimization; (3) comprehensive documentation, tutorial notebooks, a test suite with >80% tool function coverage, and integration with downstream DeepLense ML training pipelines via standard export formats (numpy, HDF5, FITS).

Mentors: Student: Prajwal Uday

Strong gravitational lensing is among the most powerful observational probes of dark matter substructure. High-resolution lensing images encode the morphological fingerprints of different dark matter models. Most available lensing images today and expected from upcoming surveys such as Euclid and LSST are ground-based and limited in resolution by atmospheric seeing. Obtaining matched high-resolution counterparts from space-based instruments like HST for every ground-based image is impractical at survey scale. This project addresses two main challenges in the proposed ML4Sci topic. First, it explores a super-resolution pipeline that does not require matched real low-resolution and high-resolution lensing observations for training. Instead, the model is trained on simulated high-resolution lensing images constructed with realistic source morphologies and survey-calibrated degradations, while remaining unsupervised with respect to real telescope HR targets. Second, it tests whether using real galaxy morphologies and limited ablations, can reduce the realism gap introduced by purely parametric Sérsic source models.

Mentors: Student: Tomoya Hatanaka

Variational quantum circuits are essential for near term quantum algorithms but designing them involves complex manual trial and error. This project solves this bottleneck by creating an automated agentic framework using Large Language Models to synthesize and optimize quantum circuits. By interacting with quantum simulators in a closed loop the AI agent will debug circuits and minimize gate depth automatically. The main deliverables include custom Orchestral AI tools for quantum simulators an iterative feedback pipeline for hyperparameter tuning and comprehensive benchmarking to uncover novel gate sequences.

Mentors: Student: ashutoshm1771

Frameworks like PennyLane and Qiskit handle circuit construction and simulation well but provide almost no tools for measuring the quantum resources a circuit uses. No unified library exists to assess profiling entanglement, magic, or Bell nonlocality. I propose a PennyLane-based library computing four resource classes through a single resource_profile() API: entanglement (Meyer–Wallach, Von Neumann entropy, concurrence), magic (Stabilizer Rényi Entropy), mana (discrete Wigner negativity for qudits), and Bell nonlocality (CHSH values). Since PennyLane QNodes are differentiable, the measures inherit gradient access for tracking resource shifts during training. I also propose benchmarking across QAOA for MaxCut, Quantum Phase Estimation, VQE for molecular ground states (H₂, LiH), and quantum autoencoders for BSM event detection following Duffy et al. (arXiv:2407.07961). The study would record resource trajectories per epoch and test whether the resource decrease Duffy et al. observed generalizes, via Spearman correlations and Mann–Kendall trend tests. This project grows out of my prior work. My Quantum Encoding Atlas library (DOI: 10.5281/zenodo.18780936) benchmarks 16 encodings across PennyLane, Qiskit, and Cirq, and my 12 QMLHEP15 evaluation tasks span quantum circuits, GNNs, equivariant QNNs, RL for PQC optimization, and contrastive learning. Running those benchmarks I identified SRE as a missing diagnostic, which inspired me towards this project. Deliverables: 1. Python package on PyPI with entanglement, magic, mana, and nonlocality modules. 2. Circuit profiler for epoch-level resource tracking. 3. Benchmarking dataset for all four algorithms. 4. Resource–performance correlation analysis. 5. API docs and Jupyter notebook tutorials. 6. Two blog posts per ML4SCI requirements. Stretch: tensor network SRE approximations beyond 10 qubits and integration with Knipfer et al.'s agent-based VQC design (arXiv:2602.19387).

Mentors: Student: Anushree Upasham

Exoplanet Atmosphere Characterisation plays a vital role in understanding chemical compositions, weather patterns and habitability of the exoplanet. Traditional retrieval models take hours per planet so using machine learning for this is a perfect alternative. I have simulated a physically realistic spectral dataset of atmospheres under regimes like equilibrium transmission, disequilibrium transmission, non-isothermal emission, and photochemical emission with instrument specific noise injection. Then two models: disequilibrium-aware GRU model and Variational Autoencoder for Pressure Temperature Profiles are used to predict various atmospheric parameters. Additionally, to detect biosignatures we can use an unsupervised convolutional Autoencoder trained on normal equilibrium spectra to flag anomalous molecular signatures. Finally, I plan on validating all my models on real JWST observational data proving that these models are worthy of real world applications.