Manu Jayadharan

Software

I treat well-engineered, open-source software as a first-class research output. All projects are on GitHub.

DAE-FINDER

github.com/mjayadharan/DAE-FINDER_dev · dae-finder on PyPI · Python

A model-agnostic scientific-ML package for discovering differential-algebraic equations from noisy data via sparse optimization. The package exposes a scikit-learn-compatible .fit()/.score() interface and supports any algorithm conforming to that API, making it a benchmarking substrate for the broader equation-discovery community (SINDy, Weak-SINDy, IDENT, and SODAs variants on a common footing). I led the design and open-source release, and I actively maintain it. Its method-agnostic architecture was inspired by the asset-agnostic pricing-library framework I worked on at Citigroup.

Recovery of algebraic relationships vs number of perturbations at 30dB SNR, small vs large perturbations
DAE-FINDER recovering power-grid algebraic relationships at 30dB signal-to-noise: larger perturbations carry more information, reaching reliable recovery with fewer experiments.

FluidLearn

github.com/mjayadharan/FluidLearn · Python, TensorFlow/Keras

A framework for solving fluid-flow PDEs using physics-informed neural networks (PINNs), packaged so that domain scientists can define geometry, boundary conditions, and physics without touching the underlying ML code.

FluidLearn architecture: a PDE operator approximated by a feed-forward neural network
FluidLearn approximates the solution and PDE operator with neural networks trained on boundary and initial-condition data.

SpaceTimeDD

github.com/mjayadharan/MMMFE-ST-DD · C++, deal.II

A parabolic-PDE solver using space-time multiscale mortar mixed finite elements with domain decomposition. Allows different spatial and temporal discretizations on different subdomains, which the time-dependent problems in the associated SIAM J. Numer. Anal. paper exploit.

Non-matching space-time subdomain grids and the computed pressure solution
Non-matching space-time grids across subdomains (left) and the computed pressure on the glued global grid (right).

BiotDDSolver

github.com/mjayadharan/BiotDD · C++, MPI, deal.II

A poroelastic fluid-flow simulator using MPI-based non-overlapping domain decomposition for the Biot system. Subdomain solves are independent and map naturally onto distributed-memory architectures.

Animated poroelastic flow simulation computed with BiotDD
A poroelastic flow simulation computed with BiotDD.

Contributions to deal.II

I am a contributor to deal.II, the widely used open-source C++ finite element library that powers most of my HPC packages.

In development