Structural analysis
Reduced structural failure by 32% using FEA simulation
How simulation-driven design optimization cut peak stress by 32% and deformation by 25% on a load-bearing industrial bracket — before any physical prototype was built.
ML-accelerated engineering
500× faster topology exploration with an ML-surrogate FEA, case study.
For an industrial-equipment client running hundreds of FEA sweeps per design cycle, Yantrix built an ML surrogate that predicts stress fields in tens of milliseconds — making full topology and parameter exploration interactive instead of overnight.
By Yantrix Engineering · ML-Accelerated Engineering2 min readIndustrial equipment
Overview
Why this study matters
A physics-informed neural network trained on 12,000 ANSYS runs replaces the full solver for early-stage topology — predicts stress fields in 40 ms vs. 22-minute solves.
Client: An industrial-equipment OEM running thousands of ANSYS solves per cycle
Project Type: Applied AI + Engineering Simulation
Industry: Industrial equipment
Service Used: ML-Accelerated Engineering + Surrogate Modeling
Results in numbers
What the engagement actually shipped.
- 500×
- Faster than ANSYS solve
- 40 ms
- Stress-field prediction
- 0.987
- R-squared on held-out geometry
- 12k
- Active-learning ANSYS runs
Objectives
What the project needed to achieve
- Predict von-Mises stress fields over parametric bracket geometry in real time
- Maintain a bounded error envelope suitable for early-stage exploration
- Integrate with the existing parametric CAD workflow
- Make the fall-back to full ANSYS solver explicit and easy
Challenge
Engineering constraint
The client’s design team was running ANSYS studies to validate bracket and frame variants. Each solve took 18-22 minutes, which capped the number of variants the team could realistically explore before freezing geometry. They asked whether ML could give them an interactive stress-field preview inside their parametric CAD workflow.
Approach
How Yantrix approached the work
- 01Defined the parameter space (thickness distribution, rib count, fillet radii, load direction) and ran an active-learning campaign to pick the 12,000 ANSYS runs that spanned it most efficiently.
- 02Trained a physics-informed neural network over stress-field outputs, using the mesh topology as a graph structure and validating against held-out physics.
- 03Deployed the model behind a FastAPI service that the SolidWorks task-pane add-in calls on every parameter change.
- 04Published a validity envelope so engineers know when the surrogate is inside its trained regime versus when to fall back to ANSYS.
Outcomes
What improved by the end
- ~40 ms inference time for a full stress-field prediction vs. 22-minute ANSYS baseline — a 500× speed-up
- R-squared of 0.987 on held-out geometries within the trained envelope
- Interactive stress preview inside the CAD tool — parameter sweeps in seconds
- Explicit validity bounds so engineers know when to defer to the full solver
Deliverables
What the client receives
- Trained PINN surrogate with documented validity envelope
- Training data pipeline reproducing the ANSYS sweep
- FastAPI model-serving container
- SolidWorks task-pane integration
- Monitoring dashboard for model drift vs. new ANSYS runs
Tools used
Stack and tooling
- PyTorch
- NVIDIA Modulus (physics-ML)
- ANSYS (for training ground truth)
- SolidWorks task-pane add-in
- FastAPI for model serving
- Weights & Biases for experiment tracking
Impact
Business-level effect
- Design-exploration throughput up by more than an order of magnitude
- Shorter convergence on final geometry → fewer full-solver validation runs
- Cultural shift: engineers explore more variants instead of guarding FEA budget
Conclusion
The real value isn’t replacing FEA — it’s making exploration cheap enough that engineers actually do it. The full solver remains the source of truth; the surrogate just makes the path there much faster.
Next step
Running a lot of similar FEA / CFD studies? Let’s talk about whether a surrogate fits — and what the training campaign would look like.
Tagged
- PINN
- ML Surrogate
- FEA
- ANSYS
- SolidWorks
- Active Learning
Frequently asked questions
Answers from the engagement itself.
Does an ML surrogate replace ANSYS?
No — it accelerates the inner loop of design exploration. ANSYS stays the source of truth for novel geometries and final certification. The surrogate predicts in milliseconds inside its trained envelope; outside the envelope it refuses to predict and engineers fall back to the full solver.
How much training data does a PINN surrogate need?
With active learning, far less than uniform sampling. For this engagement we hit 0.987 R-squared on held-out geometries with ~12,000 ANSYS runs versus an estimated 80,000+ needed for uniform sampling. The trick is in the data-selection campaign, not the model architecture.
Related case studies
Adjacent proof you can read next.
Robotics design
6-DOF robotic arm design for precise pick-and-place motion
Mechanical design of a 6-DOF compact robotic arm with cycloidal gearboxes hitting ±0.1 mm repeatability at 5 kg payload — from kinematic chain to manufactured prototype.
Continue exploring
Related blogs, services, and capability pages
Cross-links help readers move from proof into capability and educational content, and they reinforce the crawl path between commercial pages and reference content.
Let's build
Have a machine to build? Let's scope it together.
Tell us about your project. We'll respond within 1-2 business days with a preliminary scope and timeline — no boilerplate, no up-sell.