Why LLMs Predict Orbits but Fail Newton’s Laws - And What Engineers Can Do About It
Why LLMs Predict Orbits but Fail Newton’s Laws - And What Engineers Can Do About It
Why LLMs Predict Orbits but Fail Newton’s Laws - And What Engineers Can Do About It
Dr. Maor Farid, Co-Founder & CEO at Leo AI
In a recent study from Harvard and MIT, researchers demonstrated a fascinating limitation of today’s large language models (LLMs): while they can predict planetary orbits with impressive accuracy, they fail to understand the underlying physical laws driving those motions.
The experiment involved asking LLMs to predict both:
The motion (orbit, trajectory) of a planet near a central attracting object.
The attracting force (vector field) acting on that planet.
The results were surprising yet revealing. The orbit prediction was almost perfect - visually indistinguishable from the ground truth. But the predicted force field? Completely wrong.
Kepler-Level Predictions, Newton-Level Failure
This finding highlights a key point: LLMs can describe what happens (similar to Kepler’s laws of planetary motion) but cannot explain why it happens (Newton’s second law: F = ma).
In other words:
Prediction ≠ Understanding.
For engineers, this limitation is not a shock. LLMs are language models, designed to generate the most probable sequence of words. They are not inherently equipped to reason with physical laws - a core skill in engineering, physics, and mechanical system design.
Why This Matters for Mechanical Engineers
In mechanical engineering, understanding why a system behaves a certain way is crucial. Predicting stress on a beam, motion in a gear train, or airflow over a wing is not enough - the engineer must understand the physics to design safely and efficiently.
Relying solely on generic AI models for engineering decisions could lead to designs that look correct in simulation but fail in real-world application.
The Path Forward: Physics-Aware AI Models
The good news? Researchers and engineers are already developing physics-informed AI that can reason within the laws of mechanics.
Two promising approaches:
PINNs (Physically Informed Neural Networks): These integrate physical equations directly into the AI model’s training, ensuring outputs obey known laws of motion, thermodynamics, or material science.
Leo AI’s Large Mechanical Model (LMM): Designed specifically for mechanical engineers, Leo AI’s LMM understands the relationships between mechanical parts, assemblies, and physical constraints. It combines engineering datasets with embedded physics rules, enabling accurate predictions and explanations for mechanical systems.
How Leo AI Bridges the Gap
Unlike generic LLMs, Leo AI’s LMM is tailor-made for engineering. It can:
Interpret CAD and PLM data while respecting physical constraints.
Suggest design optimizations that follow Newtonian and other engineering laws.
Assist in troubleshooting mechanical systems with both predictive and explanatory insight.
This makes Leo AI not just a search or automation tool - but a physics-aware engineering assistant.
Final Thoughts: The Future of AI in Engineering
The Harvard & MIT study is a reminder: general AI is not enough for engineering. To design safe, efficient, and innovative products, mechanical engineers need domain-specific AI tools that combine data-driven prediction with physics-based understanding.
The next engineering breakthroughs will come from hybrid models - blending the creative pattern recognition of LLMs with the rigorous truth of physical laws.
Join thousands of mechanical engineers using Leo AI’s LMM to design smarter and faster.
Try Leo AI free for and experience the difference of physics-aware AI in your workflow.
Frequently Asked Questions (FAQ)
1. What did the Harvard and MIT study reveal about large language models (LLMs)?
The study showed that while LLMs can accurately predict planetary orbits, they completely fail to identify the correct force vectors acting on the planets. This highlights their inability to reason using physical laws like Newton’s second law.
2. What’s the difference between Kepler-level prediction and Newton-level understanding?
LLMs can describe what happens (like orbital motion) — similar to Kepler’s laws — but they cannot explain why it happens, which requires Newtonian reasoning (F = ma). This distinction underscores the gap between prediction and understanding.
3. Why is this limitation important for mechanical engineers?
Mechanical engineering relies not just on predicting system behavior, but on understanding the underlying physics. Without that, designs might look correct in simulation but fail in real-world applications.
4. What are Physics-Informed Neural Networks (PINNs)?
PINNs are AI models that incorporate physical laws directly into their training process. This ensures their predictions respect the rules of motion, thermodynamics, or material science.
5. How does Leo AI’s Large Mechanical Model (LMM) address this issue?
Leo AI’s LMM is specifically built for mechanical engineers. It combines engineering datasets with embedded physics rules to enable accurate predictions and physically sound explanations of mechanical systems.
6. What can Leo AI’s LMM do that generic LLMs can’t?
It can interpret CAD and PLM data while respecting physical constraints, suggest design optimizations that follow engineering laws, and assist in troubleshooting with both predictive and explanatory insights.
7. What is the future of AI in mechanical engineering?
Future breakthroughs will come from hybrid models that combine the pattern recognition power of LLMs with the accuracy of physics-based models. Mechanical engineers will need AI tools that can predict and explain using real-world laws of mechanics.
Ready to try Leo? Try Leo Today
Enjoyed this article on Why LLMs Predict Orbits but Fail Newton’s Laws - And What Engineers Can Do About It - Leo - Generative AI for Engineering CAD Design? Don’t miss Why Engineers Hate PDMs (And How AI Fixes It) - Leo - Generative AI for Engineering CAD Design
In a recent study from Harvard and MIT, researchers demonstrated a fascinating limitation of today’s large language models (LLMs): while they can predict planetary orbits with impressive accuracy, they fail to understand the underlying physical laws driving those motions.
The experiment involved asking LLMs to predict both:
The motion (orbit, trajectory) of a planet near a central attracting object.
The attracting force (vector field) acting on that planet.
The results were surprising yet revealing. The orbit prediction was almost perfect - visually indistinguishable from the ground truth. But the predicted force field? Completely wrong.
Kepler-Level Predictions, Newton-Level Failure
This finding highlights a key point: LLMs can describe what happens (similar to Kepler’s laws of planetary motion) but cannot explain why it happens (Newton’s second law: F = ma).
In other words:
Prediction ≠ Understanding.
For engineers, this limitation is not a shock. LLMs are language models, designed to generate the most probable sequence of words. They are not inherently equipped to reason with physical laws - a core skill in engineering, physics, and mechanical system design.
Why This Matters for Mechanical Engineers
In mechanical engineering, understanding why a system behaves a certain way is crucial. Predicting stress on a beam, motion in a gear train, or airflow over a wing is not enough - the engineer must understand the physics to design safely and efficiently.
Relying solely on generic AI models for engineering decisions could lead to designs that look correct in simulation but fail in real-world application.
The Path Forward: Physics-Aware AI Models
The good news? Researchers and engineers are already developing physics-informed AI that can reason within the laws of mechanics.
Two promising approaches:
PINNs (Physically Informed Neural Networks): These integrate physical equations directly into the AI model’s training, ensuring outputs obey known laws of motion, thermodynamics, or material science.
Leo AI’s Large Mechanical Model (LMM): Designed specifically for mechanical engineers, Leo AI’s LMM understands the relationships between mechanical parts, assemblies, and physical constraints. It combines engineering datasets with embedded physics rules, enabling accurate predictions and explanations for mechanical systems.
How Leo AI Bridges the Gap
Unlike generic LLMs, Leo AI’s LMM is tailor-made for engineering. It can:
Interpret CAD and PLM data while respecting physical constraints.
Suggest design optimizations that follow Newtonian and other engineering laws.
Assist in troubleshooting mechanical systems with both predictive and explanatory insight.
This makes Leo AI not just a search or automation tool - but a physics-aware engineering assistant.
Final Thoughts: The Future of AI in Engineering
The Harvard & MIT study is a reminder: general AI is not enough for engineering. To design safe, efficient, and innovative products, mechanical engineers need domain-specific AI tools that combine data-driven prediction with physics-based understanding.
The next engineering breakthroughs will come from hybrid models - blending the creative pattern recognition of LLMs with the rigorous truth of physical laws.
Join thousands of mechanical engineers using Leo AI’s LMM to design smarter and faster.
Try Leo AI free for and experience the difference of physics-aware AI in your workflow.
Frequently Asked Questions (FAQ)
1. What did the Harvard and MIT study reveal about large language models (LLMs)?
The study showed that while LLMs can accurately predict planetary orbits, they completely fail to identify the correct force vectors acting on the planets. This highlights their inability to reason using physical laws like Newton’s second law.
2. What’s the difference between Kepler-level prediction and Newton-level understanding?
LLMs can describe what happens (like orbital motion) — similar to Kepler’s laws — but they cannot explain why it happens, which requires Newtonian reasoning (F = ma). This distinction underscores the gap between prediction and understanding.
3. Why is this limitation important for mechanical engineers?
Mechanical engineering relies not just on predicting system behavior, but on understanding the underlying physics. Without that, designs might look correct in simulation but fail in real-world applications.
4. What are Physics-Informed Neural Networks (PINNs)?
PINNs are AI models that incorporate physical laws directly into their training process. This ensures their predictions respect the rules of motion, thermodynamics, or material science.
5. How does Leo AI’s Large Mechanical Model (LMM) address this issue?
Leo AI’s LMM is specifically built for mechanical engineers. It combines engineering datasets with embedded physics rules to enable accurate predictions and physically sound explanations of mechanical systems.
6. What can Leo AI’s LMM do that generic LLMs can’t?
It can interpret CAD and PLM data while respecting physical constraints, suggest design optimizations that follow engineering laws, and assist in troubleshooting with both predictive and explanatory insights.
7. What is the future of AI in mechanical engineering?
Future breakthroughs will come from hybrid models that combine the pattern recognition power of LLMs with the accuracy of physics-based models. Mechanical engineers will need AI tools that can predict and explain using real-world laws of mechanics.
Ready to try Leo? Try Leo Today
Enjoyed this article on Why LLMs Predict Orbits but Fail Newton’s Laws - And What Engineers Can Do About It - Leo - Generative AI for Engineering CAD Design? Don’t miss Why Engineers Hate PDMs (And How AI Fixes It) - Leo - Generative AI for Engineering CAD Design