Machine Learning for Engineers
Project-based AI for physical systems
An open curriculum for students and instructors who want to do more than run a model: define a physical question, build a defensible baseline, test on meaningful held-out conditions, inspect failure cases, and explain what the result means for an engineering decision.
Originating implementation: MEEG 54403: Machine Learning for Mechanical Engineers at the University of Arkansas. The course is ongoing, and this repository is designed for adaptation by other engineering programs.
Choose your path
| I am a… | Start here |
|---|---|
| Student learning a method | Read the course map, then select a module and complete its tutorial before attempting the assignment. |
| Instructor adopting one module | Use the adoption guide and the instructor guide. |
| Instructor designing AI-resilient assessment | Start with the assessment and AI policy and shared engineering-ML rubric. |
| Contributor | Read the contributing guide, especially the data-rights and student-information boundaries. |
Confirmed course implementations and adaptations are listed in IMPLEMENTATIONS.md.
What makes this different
- Engineering question first. Every module connects algorithms to a measurable physical system, variables, units, assumptions, and a decision.
- Verification before claims. Students compare against analytical, numerical, or simple empirical baselines and distinguish implementation from validation.
- Failure analysis is required. Leakage, domain shift, model-form error, label quality, uncertainty, and visibly wrong predictions are part of the work.
- AI is a tool, not evidence of learning. Students may use generative AI only under documented conditions and must verify, explain, and modify their work independently.
- Reusable by design. Python is the primary reproducible pathway; modules may be adopted independently with clear provenance and licensing expectations.
Explore the modules
- Regression: continuous engineering outcomes, baselines, residuals, and split design.
- Classification: image classification, grouped splits, and deployment constraints.
- Dimensionality reduction and clustering: structure discovery in boiling images.
- Segmentation and object detection: visual labels, transfer learning, and error audits.
- Reinforcement learning: bounded sequential decisions and policy evaluation.
- Generative models and inverse engineering design: engineering synthetic data, design candidates, and credibility checks (under development).
- Time-series forecasting and prognostics: boiling dynamics and sequence-to-sequence forecasting (validation refresh required).
- Physics-informed neural networks: transient heat conduction, inverse problems, and model-form error.
- AI agent harnesses: bounded AI-assisted engineering workflows, skills, MCP, plugins, and independent verification.
Read the curriculum architecture and roadmap for current module readiness, recommended sequencing, and planned industrial-AI extensions.
Evidence and reuse
The repository includes student project examples, a reusable data-card template, and a reproducibility checklist. Related course and project outcomes are retained as contextual evidence, not as the primary citation.
The preferred scholarly citation is Li et al. (2026). If this curriculum or adapted material informs published or other scholarly work, please cite it; the exportable metadata is in CITATION.cff. Preserve attribution and the BSD 3-Clause license, and do not upload student records, private assessment material, or restricted data.