A project-based artificial-intelligence course for mechanical engineers reported higher grade ranges over three years, but the report does not show that the teaching approach caused the change. In Fall 2021, grades were mainly between 60 and 85, with a median of about 75. They were mainly 70 to 90 in Fall 2022 and clustered in the high 80s and low 90s in Fall 2023.
The findings come from a preprint describing the University of Arkansas course known as Machine Learning for Mechanical Engineers, or MLME. It presents the course as a way to teach students how to select and implement AI algorithms to solve thermal-engineering problems. Its assessed work includes model development and implementation.
A course built around engineering problems
The course is organized as a 39.1-hour timeline covering linear regression, MLP, CNN and time-series analysis.
Its assessment totals 100 points: four assignment projects account for 50 points and a final project for the other 50. Each assignment is worth 12.5 points, with scores divided among model development, implementation, and the report or presentation at 40%, 30% and 30%, respectively.
The course was reported to have operated for three years. About 25 students were enrolled in its first year, while 14 students completed it in Fall 2023. The report does not provide complete enrollment and completion counts for every year, making the size of the overall comparison unclear.
The numbers changed alongside the teaching
The reported grade variance, a measure of how spread out scores are, was 88.02 in 2021, 42.00 in 2022 and 48.49 in 2023. That pattern suggests scores became less dispersed after the first offering, although the spread widened somewhat in the final year.
The year-to-year comparison is difficult to interpret because the course itself changed. The paper says students had no reference code in 2021 but received detailed worked examples in 2022 and 2023. Cohort differences and those instructional changes occurred at the same time as the grade shifts, so the report cannot separate their effects.
Projects extended beyond the classroom
The paper also describes student-initiated projects that produced publications involving AI model development for engineering tasks and thermal modeling or analysis for advanced device design. It does not report a publication count or establish how much of those outcomes can be attributed independently to the course.
The examples include a hole-diameter detection task in which one model, GPR, was reported to outperform another, MLP, in both detection accuracy and training speed. A cooling-system example found PCA useful for monitoring when multiple kinds of input data were used together.
Another example used images of boiling to compare four supervised models. The paper reports 100% precision for all four, while a PCA-MLP trained at 1 second per epoch and a transformer took 212 seconds per epoch. These are results for particular engineering modeling tasks, not evidence that one teaching method works better across courses or institutions.
What the report can and cannot show
The authors interpret the reported grades and project outcomes as support for project-based, scaffolded and interdisciplinary AI instruction in mechanical engineering. The report also provides a description of the course structure and examples of student work, rather than a measured comparison with another class.
The report is descriptive rather than a controlled test. It has no concurrent control group or randomized assignment, and no inferential statistical analysis is reported. The analysis covers one course at one institution, with incomplete participant counts across years; the grade distributions are approximate and are not accompanied by individual-level data.
Those limits leave open whether standard instruction would produce a similar pattern, whether the curriculum would work in larger cohorts or at other institutions, and whether any educational gains would persist into graduation, employment or job performance. The report does not establish those longer-term outcomes.
Open materials and support
The abstract states that the curriculum, syllabus, data and code are publicly available through open-access repositories. The manuscript says it was submitted to IEEE for possible publication; its acceptance and peer-review status are not reported.
The acknowledgments list support from NSF grants, University of Arkansas funds, the MathWorks Curriculum Development Support Program, ACCESS-supported computing and an Engineering Career Connection Faculty Fellowship.
Paper data and sources
Original title: Giving Mechanical Engineers Intelligent Tools: A Project-Based AI Education Curriculum in Thermal Engineering
Authors: Changgen Li, Han Hu, Christy Dunlap et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text