EngineeringNewPilot Pilot, prototype, early rollout or time-limited trial; results are limited or not yet measured.

AI predicts simulation and test results from your past data, so engineers run fewer prototypes

Added 2026-10-15 · Last updated 2026-10-15

Role

Engineering & R&D · Typical users: design engineer, simulation (CAE) analyst, test and validation engineer, R&D manager

The problem

Every design change triggers another round of finite-element or multiphysics simulation, and often another physical prototype test. Each run needs setup by a specialist and can take hours, so engineers explore only a few variants and test labs become the bottleneck. Meanwhile years of past simulation results and test reports sit in folders and are rarely reused. At Sumitomo Riko, a rubber automotive components maker, engineers must run hundreds of multiphysics simulations to understand how parts behave under extreme loads, and pre-processing such as defining geometric parameters takes extensive time and expertise.

Step by step

Pick one repeated family of analyses or tests: the same part type, the same load case, run many times a year (for example bushings, brackets, bolt joints, heat sinks).
Inventory the data you already have: number of past runs or tests, inputs recorded (geometry, material, loads, settings), outputs recorded, and whether failed or odd results were kept. A surrogate model needs dozens to hundreds of comparable runs.
Clean and label: same units, same output definitions, and a note for each run on whether it was validated against a physical test.
Train the model in the tool on most of the runs and keep 10 to 20% aside that it never sees; compare its predictions on those to the real results and record the error range per output.
Agree with the engineering manager where the prediction is good enough for screening and where full simulation or a physical test stays mandatory (safety, certification, new materials, out-of-range designs).
Use it in screening: run many variants through the model, shortlist the best, and confirm the shortlist with full simulation and the tests the tool recommends keeping.
Track for each project the simulations and physical tests run, time to design freeze, and any surprise found later in validation, and compare with similar past projects.
Retrain when you add new materials, geometry ranges or load cases; a model is only valid inside the range of data it was trained on.

Tools

Ansys SimAI: used at Sumitomo Riko
Monolith (Next Test Recommender, Anomaly Detector): used at Nissan Technical Centre Europe
Other tools to evaluate: Siemens Simcenter (Reduced Order Modeling, HEEDS), Altair physicsAI / romAI, PhysicsX, Neural Concept, and the AI features in your current CAE suite
Any LLM (ChatGPT, Claude, Copilot) with the prompt below to inventory your past simulation and test data and plan a pilot

Result

Sumitomo Riko (global maker of rubber automotive components such as vibration isolators and hoses), using Ansys SimAI trained on previously generated simulation data: AI models make performance predictions in under five minutes, over one hour of time saved for each new design, with accuracy comparable to high-fidelity simulation; initial testing showed simulation cycles over 10x faster for the mechanical performance of certain rubber bushes. Uses include anti-vibration design, battery cooling, magnetic field analysis and mixing heat transfer (Ansys press release, 8 Oct 2025). Nissan Technical Centre Europe (Cranfield, UK), using Monolith with more than 90 years of vehicle testing data: on bolt joints in vehicle chassis, the AI recommended the optimal torque range and prioritised which additional tests to run, giving an overall 17% reduction in physical testing compared with the non-AI process; Nissan says applying the approach across its European range could cut testing time by half, which is a projection, not a result (The Manufacturer, 3 Nov 2025). Both results are early or first-application results reported by the vendor or the company.

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

A surrogate model is only reliable inside the range of designs, materials and loads it was trained on; flag and fully simulate anything outside it.
Do not drop safety, certification or customer-required tests because a model predicts a pass; the customer and regulator decide what evidence counts.
Both results are early: Sumitomo Riko's 10x figure comes from initial testing on certain rubber bushes, and Nissan's 17% comes from one application (bolt joints); the 'cut testing time by half' statement is a projection.
The figures are published by the software vendors or in press coverage of their announcements; validate on a held-back set of your own runs before trusting the speed or accuracy.
Simulation and test data are core intellectual property; keep them in an approved tool and check where the vendor trains and hosts the models.

Try it yourself

You are a senior CAE and test engineer helping me plan a pilot of AI surrogate models (machine learning trained on our past simulation and test results).

I will paste: (1) a list of our past simulation runs or physical tests for one part family [run ID, date, part or variant, key inputs such as dimensions, material, load case, mesh or test settings, outputs recorded, validated by physical test yes/no, notes]; (2) the decisions these runs support [for example: pass/fail stiffness target, durability, temperature limit].

Do this:
1. Group the runs into families that use the same inputs and outputs, and list for each family the run IDs it contains (copy them; do not count or summarize numerically).
2. For each family, list the inputs and outputs that are recorded consistently and those that are missing or recorded in different units or definitions.
3. Propose which family is the best first pilot and why, based only on what is in the list.
4. Draft a validation plan: which runs to hold back for testing the model, which outputs to compare, and the questions the engineering manager must answer about acceptable error and where full simulation or physical testing stays mandatory.
5. List the safety, certification or customer-specified tests that should never be replaced by a prediction, as questions for me to confirm.

Do not calculate errors, statistics or percentages yourself; I will do them in the tool or a spreadsheet.
Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Want more like this? Join the newsletter