From Prompts to Physical Laws: A Generative AI Workflow for Engineering Physics Education
Abstract
This work explores the use of generative artificial intelligence (AI) as a source of synthetic experimental content for introductory physics courses in engineering education. Using PixVerse.ai, Grok Imagine, and Pippit, three video scenarios were generated to represent distinct resistive force regimes: constant friction, linear drag, and quadratic drag. Kinematic data were extracted from the generated videos using Tracker, an open-source video analysis tool, and subsequently fitted to the corresponding analytical models through non-linear least-squares regression in Microsoft Excel. The results show good agreement between the synthetic data and the classical kinematic equations derived from Newton's Second Law. From the fitted parameters, physically meaningful quantities were recovered in each case, with values broadly consistent with those reported in the literature under the assumed conditions. A recurring observation is that the physical plausibility of the generated motion depends on the level of detail included in the text description used to generate the videos. More specific descriptions tend to produce more coherent dynamical behaviour, suggesting that the formulation of input prompts plays a relevant role in shaping the resulting physical consistency and can be regarded as an integral component of the experimental design process. The integration of generative AI, video-based motion tracking, and curve fitting provides a complete workflow that mirrors key stages of experimental practice, from model construction to quantitative validation. The proposed methodology engages students in experimental design, data acquisition, parameter estimation, and model evaluation, promoting key engineering competencies in modelling of physical models using widely available tools. Overall, the proposed methodology demonstrates the potential of generative AI to support physics education
Related
- Development of a physical-digital learning environment for rotational motion transmission in engineering education
- CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models
- Bayesian Inversion Beyond the Generator's Range: Reducing Non-physical Constraints from Generative Neural Networks
- Engineering Quantum Interactions: From Effective Models to Validated Physical Predictions
- Artificial Intelligence Driven Physics Assignments using Context Prompts
BibTeX
@misc{alvaradocruz2026from,
title = {{From Prompts to Physical Laws: A Generative AI Workflow for Engineering Physics Education}},
author = {Laura B. Alvarado-Cruz and Josep Ll. Suñer and Pedro Yuste and Juan C. Castro-Palacio and Juan A. Monsoriu and Francisco M. Muñoz-Pérez},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.19400},
eprint = {2609.19400},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.19400},
note = {CC BY-SA 4.0},
}
RIS
TY - UNPB
TI - From Prompts to Physical Laws: A Generative AI Workflow for Engineering Physics Education
AU - Laura B. Alvarado-Cruz
AU - Josep Ll. Suñer
AU - Pedro Yuste
AU - Juan C. Castro-Palacio
AU - Juan A. Monsoriu
AU - Francisco M. Muñoz-Pérez
PY - 2026
DA - 2026/09/16
PB - arXiv
DO - 10.48550/arxiv.2609.19400
UR - https://arxiv.org/abs/2609.19400
AB - This work explores the use of generative artificial intelligence (AI) as a source of synthetic experimental content for introductory physics courses in engineering education. Using PixVerse.ai, Grok Imagine, and Pippit, three video scenarios were generated to represent distinct resistive force regimes: constant friction, linear drag, and quadratic drag. Kinematic data were extracted from the generated videos using Tracker, an open-source video analysis tool, and subsequently fitted to the corresponding analytical models through non-linear least-squares regression in Microsoft Excel. The results show good agreement between the synthetic data and the classical kinematic equations derived from Newton's Second Law. From the fitted parameters, physically meaningful quantities were recovered in each case, with values broadly consistent with those reported in the literature under the assumed conditions. A recurring observation is that the physical plausibility of the generated motion depends on the level of detail included in the text description used to generate the videos. More specific descriptions tend to produce more coherent dynamical behaviour, suggesting that the formulation of input prompts plays a relevant role in shaping the resulting physical consistency and can be regarded as an integral component of the experimental design process. The integration of generative AI, video-based motion tracking, and curve fitting provides a complete workflow that mirrors key stages of experimental practice, from model construction to quantitative validation. The proposed methodology engages students in experimental design, data acquisition, parameter estimation, and model evaluation, promoting key engineering competencies in modelling of physical models using widely available tools. Overall, the proposed methodology demonstrates the potential of generative AI to support physics education
ER -