AI × ENGINEERING
Explore AI in Engineering
Engineers use AI to inspect images, explore candidate designs and approximate simulations. The result has to be evaluated against requirements such as fit, load, cost or acceptable error.
Examine a NASA hardware-design case, learn how engineers check model outputs, and compare two projects with measurable requirements.
Select a step to see what happens.
Requirements
Specify what the part must do
Define attachment points, loads, available space, and a measurable objective before comparing candidate designs.
Design options
Explore the allowed design space
A generative design method can propose structures within the supplied constraints. A striking shape is only a candidate.
Engineering review
Check the proposed part
Review fit, stress, material behavior, and manufacturability. Simulation and a generated drawing do not establish physical performance.
Schematic design workflow. The drawing is not a validated component.
THE BASICS
Start with a requirement you can measure
Suppose a bracket must hold a sensor in a particular position. Making it lighter is useful only if it still carries the required load, fits its neighbors, and can be made. That is an engineering problem: choose a design while respecting constraints. The requirements determine which candidate designs are worth considering.
Different methods suit different engineering tasks. Computer vision can identify unusual image patterns. Optimization searches for designs that meet constraints while improving an objective, such as material use. A learned simulation model approximates results from numerical examples. An agent can organize tools, calculations and documents within one of these workflows.[2][1][3]
Name the input, output, unit, and comparison. Ask what a simpler approach would do and which mistakes matter most. These questions are useful before you have advanced mathematics or a workshop.
APPLICATIONS
Where can AI support an engineering decision?
Follow the information through each task: the input, the model’s contribution, and the person responsible for checking the result.
Public industrial-inspection research benchmark
Finding image patterns that need inspection
An inspection system needs to distinguish a relevant defect from normal variation in an object or texture.
- The material
- The MVTec AD benchmark includes object and texture images: defect-free training examples, test examples with and without defects, and marked defect regions.
- AI’s role
- An anomaly-detection method learns what usual examples look like and produces a score or candidate region for further inspection.
- The human decision
- Define the defects that matter, choose a decision threshold and inspect both missed defects and unnecessary alerts.
The limitA change in camera, lighting, material or production conditions can change the evaluation. A benchmark result alone does not establish performance on a particular production line.[2]
Documented NASA hardware-design work
Exploring shapes inside a design boundary
A component must connect particular surfaces while leaving room for other equipment and assembly.
- The material
- The design process starts with connection geometry, load requirements, and spaces that must remain empty.
- AI’s role
- NASA describes generative design producing candidate structures within those restrictions.
- The human decision
- Engineers define the objective, review candidate shapes and carry out the analysis needed to judge whether a part is practical.
The limitThe generated shape still needs engineering analysis and validation before it can be treated as an approved component.[1]
Published computational research
Learning an approximation to a simulation
Exploring a system can require many calculations as starting conditions change.
- The material
- Numerical examples link inputs to solutions of a specified physical model.
- AI’s role
- Fourier Neural Operator research learns relationships between input functions and solution functions. Its experiments included fluid-flow and other differential-equation problems.
- The human decision
- Compare predictions with reference solutions on withheld cases and inspect where error grows. Decide whether the approximation suits the intended analysis.
The limitA successful numerical benchmark does not validate an unfamiliar physical system. A simulated result also depends on the equations and assumptions used to generate it.[3]
CASE STUDY
How NASA defines and checks a generated part
Attachment points, loads and clearance requirements define the design problem before generation begins.
NASA's February 2023 account describes a CAD specialist specifying attachment surfaces and keeping clear paths for equipment, light, or assembly access. The algorithm then proposes a structure connecting the allowed areas. Those requirements determine which structures the algorithm can propose.[1]
The account also notes that an algorithm can produce sections that are too thin. Engineers review the output, and parts undergo NASA's established analysis and validation processes. Generation is followed by review and validation.[1]
To compare candidate parts, report estimated mass alongside the requirements for fit, stress, assembly and material behavior. The checks depend on the component and its intended use. A design explanation should make those requirements and the remaining validation work explicit.
Follow the process Open a step to see what happens.
Define
Connections, loads, and spaces that must stay clear.
Explore
Candidate structures within those boundaries.
Evaluate
Human review and appropriate engineering analysis.
A useful design report explains the requirements, the candidate structure and how it will be checked.[1]
SKILLS & STUDY
What to learn first
- Quantities, units, and assumptions
- Keep mass distinct from force, and energy distinct from power. Show unit conversions and state which conditions are fixed before comparing results.
- Baselines and measurements
- Define a simple comparison, a measurable outcome, and an acceptable error. Include unexpected results and representative errors so another reader can assess the comparison.
- Domain foundations
- Geometry and algebra support early exploration. Mechanics, circuits, materials, calculus, and numerical methods become relevant as the physical question becomes more demanding.
- Evaluation and communication
- Describe the intended use, missing evidence, and consequences of errors. NIST's voluntary AI RMF offers a framework for considering risk across a system's life cycle.
PROJECT IDEAS
Projects built around measurable constraints
Choose a question, inspect the requirements, and keep a brief for later. Each proposal includes a way to check the work.
Save an idea in this browser or copy its brief to discuss with a mentor.
No account neededWhat does an inspection score miss?
For one public image category, which kinds of defects would a selected detector miss or flag unnecessarily?
- Intended output
- An evaluation report with a defined image split, a simple baseline, example errors, and an explanation of the chosen threshold.
Skills, sources and checks
- Starting knowledge
- Python arrays, image basics, and the distinction between training and testing. If you are still learning these foundations, begin by interpreting the dataset’s documented examples and results.
- Material to consider
- One MVTec AD category, its annotations, and the dataset's stated license and evaluation guidance.
- How to judge the result
- Keep test images separate from model selection. Report missed defects and false alerts alongside the overall score.
Evaluate the selected benchmark category and describe which other conditions would need separate testing.
A parts plan with explicit trade-offs
Which combinations of fictional components satisfy a small set of stated size, cost, and availability constraints?
- Intended output
- A reproducible comparison table and a short recommendation explaining which requirements drive the decision.
Skills, sources and checks
- Starting knowledge
- Spreadsheets, inequalities, and checking combinations. A later software version can compare a search method with an exhaustive baseline.
- Material to consider
- A clearly labeled synthetic parts table, with invented dimensions, prices, and quantities; keep the same units throughout.
- How to judge the result
- Check every constraint and recompute totals independently. Change one assumption and explain when the preferred combination changes.
A planning comparison using fictional specifications. Evaluating an assembled device would require verified components and physical checks.
These ideas investigate evidence and models. Any physical build needs suitable supervision, verified specifications, and checks appropriate to the equipment and intended use.
MENTOR SUPPORT
Scope an engineering project with a mentor
The STEMtiq Agentic AI Scholars Program supports an agent-building project with mentor guidance. An engineering interest could lead to a research, calculation or planning workflow, subject to a mentor’s assessment of the question and your preparation.
Discuss the data, baseline and evaluation before choosing a direction. The shared curriculum covers agent-building; it does not include a dedicated robotics, hardware, CAD or simulation course.
5 weeks · 25 live hours · Grades 9-12
COMMON QUESTIONS
What students ask
Does AI design a finished engineering solution?
It can propose options or approximate an analysis. Someone still defines requirements, checks the evidence, and decides which additional tests are needed. NASA's example includes human inspection and established validation after generation.[1]
Can I explore this before taking calculus?
Yes. You can compare quantities, read a design case or investigate a planning table using algebra and spreadsheets. Numerical simulation and advanced model training require additional mathematics.
Why isn't a simulation proof that a design works?
A simulation describes behavior inside a model with selected assumptions and inputs. Check whether those assumptions represent the situation you care about, then identify the independent measurements or physical tests needed for that claim.
What should an AI inspection evaluation report?
Describe the images, held-out cases, decision threshold, missed defects, and unnecessary alerts. Include error examples. The MVTec benchmark offers annotated images for research, but another setting needs its own evaluation.[2]
Does following an AI framework certify my project?
No. NIST AI RMF 1.0 is voluntary risk-management guidance. It can help organize questions about context, measurement, and consequences; using its vocabulary does not certify a model or replace engineering validation.[4]
Define the requirements for your project
Write down the input, expected output, constraints and comparison you would use. A mentor can help assess whether that plan fits the program’s software-and-evidence project format.
Sources & further reading
- [1]
- [2]MVTec Software: MVTec AD industrial anomaly detection dataset (Dataset introduced in 2019; current resource checked October 2026)Return to guide ↑
- [3]
- [4]
Sources inform this guide; their organizations do not endorse STEMtiq. Project ideas and diagrams are original editorial explanations.