[Copy] Lockheed Martin accelerates design with AI and embedded simulation
Lockheed Martin collaborated with nTop to rethink how complex systems are designed. By embedding simulation directly into the modeling workflow, eliminating fragile meshing, and using AI to drive optimization, the team evaluated over 400 heat exchanger designs in under eight hours.
The result: a process that can shrink design iteration cycles from weeks to minutes.
- Organization: Lockheed Martin
- Industry: Aerospace & Defense
- Application: Heat Exchanger Optimization
- Tools Used: nTop + NVIDIA Modulus
The project
Redefining thermal-fluid design, and building AI-ready workflows
Lockheed Martin set out to optimize the design of a dual-fluid heat exchanger featuring a gyroid-based TPMS core. The engineering goal was straightforward but technically challenging: maximize heat transfer while minimizing pressure drop all under varied inlet flow conditions and within additive manufacturing constraints.
The challenge
When traditional CFD stalls innovation and slows institutional learning
Legacy simulation workflows are ill-suited for high-speed design iteration. Meshing failures, long solve times, and fragile geometry slow progress dramatically. This was especially true when working with advanced geometries like TPMS, where even setting up a single analysis could take weeks.
Beyond speed, these brittle workflows make it hard to capture and reuse hard-earned setup knowledge, meaning each engineer has to rediscover solutions to the same problems. With emerging AI-ready workflows, that knowledge becomes increasingly reusable.
The solution
Embedded simulation meets AI-driven inverse design
The nTop team developed a parameterized implicit model of the heat exchanger and used voxel-based CFD to eliminate the need for meshing entirely. A voxel convergence study ensured fidelity, while a Latin Hypercube sampling strategy enabled broad design space coverage. Over 400 design points were simulated in just 6–8 hours with zero failures.
Data Generation Pipeline
This dataset then powered two machine learning models: a neural network for predicting scalar quantities like pressure drop and heat transfer, and a Fourier Neural Operator (FNO) for full-field predictions. These models showed early promise in supporting real-time optimization loops, enabling initial demonstrations of inverse design in seconds.
Fourier Neural Operator - CFD Surrogate Model
The results
400+
0
meshing or solve failures
Milliseconds
to predict flow and pressure fields using surrogate models
Seconds
to complete inverse design optimizations
Weeks to minutes
in potential cycle time reduction
An in-progress, AI-ready workflow
An in-progress, AI-ready workflow
that Lockheed aims to apply to other thermal-fluid and structural systems
Results: FNO Prediction - Velocity X (on training sample)
Left - Inference time: 1 second | Right - Simulation time: 430 seconds
Results: FNO Prediction - Velocity X (on validation sample)
Left - Inference time: 1 second / Right - Simulation time: 3600 seconds
Results: FNO Prediction - Pressure (on training sample)
Left - Inference time: 1 second / Right - Simulation time: 430 seconds
Results: FNO Prediction - Pressure (on validation sample)
Left - Inference time: 1 second / Right - Simulation time: 3600 seconds
Why nTop?
Implicit modeling for parameterized design
Engineers at nTop created a fully parametric heat exchanger model using implicit geometry. This enabled high design flexibility while maintaining geometric robustness under rapid iteration – a requirement for AI-driven iteration, where dozens of changes can happen in seconds.
No meshing, no crashes, no rework
With voxel-based simulation, the workflow eliminated the need for meshing entirely, which is a major hurdle for TPMS structures. This meant no meshing failures, no rebuilds, and no iteration downtime, and gave AI models clean, consistent geometry to reason with.
Accurate, scalable CFD
nTop provided fast, reliable simulations that captured nonlinear fluid interactions, including transitional turbulent zones, to accurately predict pressure drop across the heat exchanger.. A voxel resolution study helped strike the right balance between fidelity and speed.
Machine learning integration
Using simulation data, Lockheed engineers trained neural network and FNO models capable of both scalar and full-field predictions. These models provided high accuracy with millisecond evaluation times ideal for optimization.
Built to support Lockheed’s AI infrastructure goals
The workflow is being evaluated for integration into Lockheed Martin’s broader AI design infrastructure, with the goal of enabling scalable application to other complex systems. This early success suggests that the path to AI-enabled engineering can be accelerated by building robust, reusable workflows today.
Conclusion
Lockheed Martin and nTop are exploring the potential of simulation-driven, AI-enabled engineering. By embedding simulation into modeling and combining it with intelligent surrogate modeling, the team has made early progress toward proving that engineering teams can move faster, explore more, and trust the results.
The next step? Continuing to mature and scale this approach across Lockheed’s high-impact product lines