An Agentic AI Workflow to Simplify Parameter Estimation of Complex Differential Equation Systems

Saakaar Bhatnagar · 2025

Read An Agentic AI Workflow to Simplify Parameter Estimation of Complex Differential Equation Systems by Saakaar Bhatnagar (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-07283.txt.

Introduction

Saakaar Bhatnagar, writing alone in 2025, presents an AI agent workflow that automates the fitting of differential equation models. Ordinary differential equations describe how quantities change over time in fields from biology to engineering. Fitting their parameters to data is slow manual work: data are noisy and partial, models can be stiff, and coding a fast version demands special skill.

The user supplies a light XML description of the problem and fills in a code skeleton. The agent checks that the two agree, fixes flaws in the input, and converts the code into JAX functions, a form that compiles and runs in parallel. A two-stage search then scans broadly before refining with gradients. The open-source release aims to move users from a problem statement to an interpretable fitted model with far less effort, while experts keep control.

Catalog record

CREATORSaakaar Bhatnagar
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-07283

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

An Agentic AI Workflow to Simplify Parameter Estimation of Complex Differential Equation Systems Saakaar Bhatnagar Abstract Parameter identification for mechanistic Ordinary Differential Equation (ODE) models underpins prediction and control in several applications, yet remains a manual and labor-intensive process: datasets are noisy and partial, models can be stiff and complex, and differentiable implementations demand framework expertise. An agentic AI workflow is presented that converts a lightweight, human-readable specification into a compiled, parallel, and differentiable model calibration pipeline. Users supply an XML description of the problem and fill in a Python code skeleton; the agent automatically validates consistency between problem definition and code, and auto-corrects pathologies in the input deck. It transforms Python callables into pure JAX functions for efficient just-in-time compilation and parallelization. The system then orchestrates a two-stage search comprising global exploration of the parameter space followed by gradient-based refinement. The result is an AD-native, reproducible workflow that lowers the barrier to advanced calibration while preserving expert control. An open-source implementation with a documented API and examples is released, enabling rapid movement from problem statement to interpretable ODE models with minimal effort. Keywords: Agentic AI; Parameter Estimation; Differentiable Physics; Ordinary Differential Equations (ODEs); JAX 1 Introduction Ordinary differential equation (ODE) parameter fitting is a foundational task for turning mechanistic models or ideas into predictive tools across the sciences and engineering. In battery models, ODEs govern degradation kinetics, thermal behavior, and equivalent-circuit dynamics [ 1 , 2 ] ; accurate parameters are essential for state estimation, control, and safety analysis. In combustion modeling, reactions are often represented using Arrhenius ODE systems [ 3 , 4 ] . In biology, systems of ODEs describe gene regulation, signaling, and epidemiological dynamics [ 5 ] . More broadly, ODE systems are frequently embedded within larger multiphysics simulations [ 2 , 6 ] and digital twins, and poor calibration at the ODE level propagates downstream, degrading overall fidelity, stability, and decision quality. However, in many instances, the parameters of the ODE models need to be estimated from data. Rigorous estimation of these parameters remains difficult in practice: real-world datasets are noisy and partial; models are often stiff or multiscale; and calibration requires repeated numerical solves which are computationally intensive and fragile to implementation details. A variety of strategies exist to calibrate these ODE models to data. Problem-specific heuristic methods, where techniques such as equation linearization reduce the estimation problem to a linear regression [ 7 , 8 , 9 ] , have been demonstrated in several use cases, but suffer from robustness and generalizability issues [ 10 ] . Population-based algorithms (like genetic algorithms or swarm optimization) have been successfully used in applications such as battery modeling [ 11 , 12 ] , structural engineering [ 13 ] , and piezoelectrics [ 14 ] . However, population-based methods suffer from poor convergence properties near local minima, and scale poorly with search space dimension. This has been addressed in the past using gradient-based search methods near local optima. In particular, derivatives obtained using automatic differentiation instead of numerical differences have shown considerable success in solving ODE and PDE (Partial Differential Equation) model identification problems [ 15 , 16 , 17 ] , particularly when models exhibit stiffness characteristics [ 10 , 18 ] . By backpropagating through the ODE solve via algorithmic differentiation of the integrator or adjoint sensitivity methods, exact derivatives of the time-discretized simulation [ 19 ] output are obtained with respect to parameters, compared to numerical differencing which has inherent discretization errors. However, practical adoption of efficient model identification algorithms is hindered by substantial software and workflow barriers. In particular, writing fully differentiable simulation code requires nontrivial fluency in state-of-the-art AD frameworks (e.g., PyTorch, JAX, Julia) and with their distinct programming models. For example, JAX’s functional, transformation-driven style (pure functions with explicit jit /vmap/grad semantics) can be unfamiliar for domain scientists. Beyond framework fluency, practitioners must hand-engineer complete optimization pipelines—defining objectives and constraints, organizing multi-experiment datasets, selecting solvers and tolerances, configuring sensitivity modes, and managing compilation and hardware particulars—each step demanding bespoke code and iterative tuning. These burdens reduce accessibility to modern capabilities such as reverse-mode automatic differentiation, just-in-time (jit) compilation, and advanced optimisers. This motivates tooling that abstracts these concerns and automates routine choices while preserving expert control. Recent advances in artificial intelligence have delivered substantial improvements in user workflow efficiency, and several studies have investigated using LLMs to improve workflows in scientific computing. Kashefi and Mukerji 2023 investigated the use of ChatGPT for writing and debugging numerical algorithms for a variety of complex problems. Dong et al. 2025 and Pandey et al. 2025 investigated using Large Language Models (LLMs) to automate the setup and running of complex fluid dynamics simulations, reducing the need for domain expertise and leading to smoother workflows. Further work has also explored using autonomous AI Agents to automate and accelerate engineering. Elrefaie et al. 2025 propose a multi-agent framework for aerodynamic design, streamlining key steps in the automotive design process such as concept sketching, shape retrieval, meshing and simulation. Picard et al. 2025 explored using Vision-Language Models (VLMs) in tasks relevant to engineering design such as similarity analysis, CAD generation, and topology optimization. These works demonstrated the potentially large impact AI agents and workflows could have in improving the efficiency of engineers and designers developing new products, similar to the already-realized productivity gains made in fields like software engineering and data analytics [ 25 ] . In this work, the use of agentic AI workflows is investigated to streamline parameter identification for complex ODE systems. The proposed framework abstracts away much of the human effort and multidisciplinary expertise traditionally required to formulate and solve these problems while remaining compatible with modern differentiable programming stacks. By delegating routine orchestration to an AI model and exposing only high-level controls to the user, the proposed workflow lowers the barrier to employing just-in-time compilation, automatic differentiation, and advanced optimization methods for ODE calibration. The result is a more accessible, reproducible, and efficient path from problem statement to calibrated model, without sacrificing numerical rigor or expert oversight. The framework is demonstrated on two common types of parameter estimation problems: recovering the parameters from simulated data, and obtaining models that best describe experimentally obtained data. This work has the following features that would make it a valuable contribution: 1. Simplified usage interface : A human-readable XML interface and standard Python functions are used to define the problem. These are automatically translated into a well-posed constrained optimization in a few steps; capturing objectives, bounds, and constraints—thereby eliminating boilerplate scripting and reducing specification errors. 2. AD-native, jit -compiled execution : Standard Python callables are automatically transformed into pure JAX functions and staged for just-in-time compilation, enabling stable reverse-mode automatic differentiation and parallel execution (e.g., via pmap) without exposing users to framework-specific plumbing. This abstraction delivers differentiable, high-performance kernels from ordinary Python code. 3. Automated validation of optimization setups : The system performs automatic checks prior to fitting, identifying common pathologies such as infeasible or ill-conditioned inputs, misconfigured objective/constraint functions, and parameters declared but unused. The agentic setup automatically corrects these issues, and informs users of further improvements they can make before launching the fitting problem. Early diagnostics eliminate incorrect parameter estimation runs, reducing cost.

Read the complete text