From Prompts to Physical Laws: A Generative AI Workflow for Engineering Physics Education

Synthesis: Alvarado-Cruz et al. (2026) test whether generative AI video can supply the experimental material for introductory engineering physics. Using PixVerse, Grok Imagine and Pippit, they generated three scenarios representing distinct resistive-force regimes — constant friction, linear drag and quadratic drag — then extracted kinematic data with Tracker and fitted analytical models by non-linear least squares in Excel. The synthetic data agreed well with the classical equations of motion, and physically meaningful parameters were recovered. The recurring practical finding is that the physical plausibility of the generated motion depends on how specifically the prompt describes the situation, making prompt formulation part of experimental design.

Key Findings

1. Generated video can serve as synthetic experimental content. Kinematic data extracted from AI-generated clips fitted the expected models for constant friction, linear drag and quadratic drag.
2. Fitted parameters recovered physically meaningful quantities broadly consistent with literature values under the assumed conditions.
3. Prompt specificity governs physical coherence. More detailed descriptions produced more coherent dynamical behavior, so prompt formulation becomes an explicit stage of the experimental design process rather than a preliminary convenience.
4. The workflow mirrors real experimental practice end to end — model construction, data acquisition, parameter estimation and model evaluation — using an Open Source tracker and spreadsheet regression.
5. The pedagogical aim is competence, not convenience: students engage in experimental design and quantitative validation rather than watching a demonstration.

What the workflow offers when laboratory access is constrained

The value proposition is not that AI video replaces a laboratory; it is that it provides varying physical regimes on demand, including ones a teaching lab cannot easily produce. That matters for Inquiry-Based Learning where the learner must decide what to measure and which model applies, and it is why the paper frames the activity as Experiential Learning rather than as a demonstration. It also sidesteps a real constraint in physics education: apparatus cost and scheduling.

The validity question the design must answer

Synthetic data that merely looks plausible would teach students to fit models to artifacts. The authors' checks — agreement with Newton's second law and plausible recovered parameters — are the right first tests, but they are the authors' own validation, not a student learning outcome. Read alongside the wiki's work on Simulation and Assessment Validity, the paper defines a workflow that still needs evidence on whether students learn measurement judgment from it, and it inherits the general reliability concerns that attach to generated content used as data.

What this means for practice

• Instructors. Make prompt specificity a graded stage of the experiment: have students write the scene description, generate the clip, and defend why that description should produce the intended resistive regime before any data are extracted.
• Instructors. Run the four phases end to end with students — generation, center-of-mass tracking in Tracker, non-linear least-squares fitting in Excel or Python, then interpretation of the recovered parameter against literature values — rather than showing a finished clip.
• Instructors. Assign the regimes a teaching laboratory cannot easily produce (glycerin-viscosity linear drag and an 80 kg skydiver's quadratic drag alongside constant friction) so that students must decide which model fits each clip.
• Administrators. Fund AI video generation as a supplement to constrained laboratory access, not a replacement for apparatus-based measurement, since the paper demonstrates author-side model fits and reports no student learning outcome.

Limitations

• The demonstration rests on three generated clips, one per resistive regime, each from a different platform (PixVerse.ai, Grok Imagine, Pippit); no replications across prompts or models are reported, so the prompt-specificity conclusion comes from this small set of generations.
• No students took part. The paper reports fit statistics (R² = 0.9982 for constant friction, 0.9967 for linear drag, 0.9953 for quadratic drag) rather than any learning, engagement or measurement-judgment outcome, so its competence claims remain design rationale.
• Validation is the authors' own fit of a reparametrized model to tracked data; a near-unity R² establishes that the clip follows the assumed equation, not that the physics is correct, since a biased clip can still be fitted by the same functional form.
• Recovered coefficients depend on assumed masses and dimensions (silicone sphere at ρ ≈ 1200 kg/m³ and d = 0.05 m, skydiver m = 80 kg), and no uncertainty analysis or propagation of tracking error is reported.

Connected Concepts

• Generative AI
• Simulation
• Physics Education
• Experiential Learning
• Inquiry-Based Learning
• Prompt Engineering
• Multimodal AI
• Active Learning
• Assessment Validity
• Hallucination Risk
• Video in Education

Connected Articles

• It's Not the Tool, It's the Task: A Framework for Cognitively Activated AI Augmentation in Physics Instruction — Cognitively activated AI in physics
• Large Scale AI Grading of Handwritten Physics Assessments: Score Agreement and Olympiad Team Selection Outcomes — AI grading of handwritten physics work
• AI in Particle Physics Education: Research Problems and Foundational Skills — Redesigning particle physics education with AI
• AI-based scoring systematically underestimates conceptual understanding of linguistically weak students' explanations in physics — Language bias in AI scoring of physics responses

Citation

Alvarado-Cruz, L. B., Suñer, J. Ll., Yuste, P., Castro-Palacio, J. C., Monsoriu, J. A., & Muñoz-Pérez, F. M. (2026). From Prompts to Physical Laws: A Generative AI Workflow for Engineering Physics Education. arXiv:2609.19400.