Advisor - Scientific Machine Learning & Agentic Workflows Engineer

Eli Lilly and Company • Us, IN

3d • Onsite

About The Position

This role builds hybrid physics-and-data models — and the agentic software layer that puts them in the hands of working scientists. It has two connected halves. The first is scientific machine learning: physics-informed networks, operator learning, multi-fidelity surrogates, Bayesian calibration, and gray-box system identification that accelerate or extend physics-based simulation. The second is AI engineering: agentic workflows that plan, set up, execute, and post-process modeling and simulation tasks by calling real solvers and real data, so that a scientist can move from question to credible answer without hand-assembling every step. The role sits within the Computational Modeling & Simulation team in DDCS and works across the programs that the team supports. This is not a standalone research role: the models and tools are built with and for the drug product, device, and process development functions across Product Research & Development that use them. You will independently design, implement, validate, and support the workflows you build, working closely with the DDCS AI Application Development and Data Sciences functions on architecture, platform choices, and compliance.

Requirements

• PhD in applied mathematics, computer science, machine learning, computational mechanics or physics, chemical, mechanical, or biomedical engineering, or a related field
• Doctoral research centered on scientific machine learning — physics-informed learning, operator learning, surrogate modeling, or hybrid mechanistic-ML methods.
• Demonstrated experience developing SciML models and validating them against physics-based simulation or experimental data.
• Strong Python software engineering, including PyTorch or JAX, and Git-based collaborative development.
• Hands-on experience building LLM-enabled workflows including multi-step agent orchestration and either tool and function calling or retrieval-augmented generation (RAG).
• Experience running computational work on HPC clusters or cloud infrastructure.
• Peer-reviewed publications or open-source contributions in scientific machine learning.

Nice To Haves

• A working practice grounded in evaluation, provenance, and reproducibility, and the ability to explain to both technical and business audiences what these methods can and cannot do.
• Neural operator methods, multifidelity modeling, Bayesian UQ and calibration, active learning, or Bayesian optimization applied to engineering or biomedical problems.
• Gray-box identification, symbolic regression, or discovery of pharmacokinetic, pharmacodynamic, or transport models from sparse data.
• Experience with agent frameworks and tool-interoperability standards (e.g., tool-calling APIs, graph-based agent orchestration, Model Context Protocol) and with production deployment practice (containers, CI/CD, observability).
• Familiarity with commercial or open-source simulation solvers and their scripting interfaces and input/output formats.
• Awareness of model credibility practice for regulated products (ASME V&V 40; FDA credibility assessment guidance) and of GxP and 21 CFR Part 11 requirements for computerized systems.
• Front-end or full-stack experience sufficient to build usable interfaces for scientific tools.
• Domain exposure to drug delivery, medical devices, combination products, or biomedical transport problems.

Responsibilities

• Independently design, train, and validate SciML models — physics-informed neural networks, operator-learning architectures (e.g., DeepONet, Fourier neural operators), and Gaussian-process or multifidelity surrogates — against high-fidelity simulation and experimental data.
• Apply Bayesian calibration and uncertainty quantification to deliver predictions with a defensible confidence statement rather than a point estimate.
• Apply gray-box identification and symbolic-regression methods to infer unknown parameters or missing mechanisms from sparse experimental data.
• Design and build LLM-based agentic workflows that plan, set up, execute, monitor, and post-process modeling tasks by calling real tools — solvers (e.g., Abaqus, COMSOL, Ansys, OpenFOAM, LAMMPS, GROMACS), meshing and geometry utilities, HPC schedulers, and internal data services.
• Implement the tool interfaces, APIs, and retrieval layers that connect agents to DDCS model libraries, simulation archives, and structured data sources.
• Define and enforce human-in-the-loop checkpoints at the points where a modeling decision requires expert judgment rather than automation.
• Package workflows so that a scientist who is not a software developer can use them reliably and unaided.
• Ensure every agent-executed run emits a complete provenance record: inputs, geometry and discretization, solver and library versions, convergence evidence, random seeds, and the human approvals applied.
• Design for deterministic replay — any result that informs a decision must be reproducible from its recorded provenance, within a documented tolerance.
• Build evaluation harnesses — benchmark problems with known solutions, regression tests, and reliability metrics — that quantify how often a workflow produces the right answer.
• Define credibility practice: defining defensible practice and credibility frameworks for hybrid physics-ML models in the spirit of ASME V&V 40 and the FDA's 2023 guidance on assessing computational model credibility.
• Write maintainable, tested code; treat version control, code review, CI/CD, and containerization as defaults.
• Deploy and operate workflows on HPCs and approved cloud environments, instrumented for observability, latency, and cost.
• Deploy in partnership with the functions that will use the output — device engineering, drug product and process development, analytical sciences, manufacturing, quality, and regulatory — so that tools fit existing development workflows, data sources, and decision timelines rather than requiring users to change how they work.
• Apply business judgment when setting priorities: understand the portfolio, program milestones, and decision gates the modeling supports, and direct effort toward the questions where a faster or better answer changes a development decision.
• Make and defend practical trade-offs on build versus buy, model fidelity versus cost and turnaround time, and automation versus expert review, keeping total cost of ownership and the needs of downstream users in view.
• Apply responsible AI, security, and data-handling controls across the workflow lifecycle, in partnership with IT, Quality, and Information Security.
• Work as an embedded member of cross-functional development teams, engaging directly with drug product, device, and process development colleagues to frame the problem and agree what a useful answer looks like before choosing a method.
• Partner with continuum and molecular modeling scientists to identify where surrogates and agentic workflows create real value — and where a direct simulation or an experiment remains the better answer.
• Collaborate with the DDCS Digital Transformation team on architecture, reusable patterns, and platform choices; contribute to GxP and 21 CFR Part 11 considerations where workflows touch regulated systems.
• Train and support users, and report adoption quantitatively.

Benefits

• company bonus (depending, in part, on company and individual performance)
• company-sponsored 401(k)
• pension
• vacation benefits
• medical, dental, vision and prescription drug benefits
• flexible benefits (e.g., healthcare and/or dependent day care flexible spending accounts)
• life insurance and death benefits
• certain time off and leave of absence benefits
• well-being benefits (e.g., employee assistance program, fitness benefits, and employee clubs and activities)

What This Job Offers

• Job Type: Full-time
• Career Level: Principal
• Education Level: Ph.D. or professional degree