Can AI Learn Physics from Sensor Data? Archetype AI’s Breakthrough

The physical world is complex and often chaotic, governed by the laws of mechanics, thermodynamics, and other natural phenomena. Humans have long relied on observation and precise measurement to uncover these laws. But could AI achieve the same insights by analyzing sensor data without human guidance? What if AI could help us understand the behaviors of complex systems, such as electrical grids, industrial machinery, or human-computer interactions, in ways that elude traditional equations?

A Milestone for Archetype AI’s Physical Foundation Model

Archetype AI Team is proud to announce a major step forward in developing a physical AI foundation model. In a recently published paper, "A Phenomenological AI Foundation Model for Physical Signals," the team demonstrated how an AI model can encode and predict physical behaviors it has never encountered before, without being explicitly taught the underlying laws of physics. These findings could revolutionize how AI interacts with and understands the physical world.

Rethinking How AI Learns Physics

The traditional assumption is that AI, like humans, must first learn the laws of physics before applying them to real-world tasks. This method introduces inductive biases—assumptions that constrain the AI's knowledge within predefined mathematical principles. However, training AI in this way often results in highly specialized models that struggle to generalize to different systems or scenarios.

For instance, a model trained to analyze fluid motion using the Navier–Stokes equations would be ill-equipped to interpret radar Doppler images, which involves different physical principles. Similarly, complex systems like electrical grids or industrial machinery defy simple physical laws, making it difficult to describe them with basic equations.

A New Approach: Letting AI Discover the Laws of Physics

What if we let AI discover the laws of the physical world by analyzing raw sensor data, much like Johannes Kepler and Georg Ohm uncovered the laws of planetary motion and electricity through observation and measurement? Instead of teaching AI specific principles, Archetype AI aims to have its model, Newton, find hidden patterns and statistical distributions in the data.

The difficulty in comprehending the physical world, for both humans and AI, lies in the fact that we can’t directly experience it—we rely on sensors to observe it indirectly. These sensors range from our natural biological senses, like sight and hearing, to a wide array of artificial sensors designed to measure everything from speed to gas levels. However, sensors inevitably influence the results, distorting or masking the 'true' physical behaviors, which complicates the task of uncovering the fundamental laws that govern these systems

The true test of whether an AI model understands physical systems is its ability to generalize beyond the data it was trained on. For example, could a model trained on data from a car's electric motor predict noise pollution in a distant city? Or could a model analyzing a butterfly's wing movement forecast hurricane dynamics across the globe? This allows the model to accurately characterize physical systems it has never previously encountered, even when those systems are measured using sensors it hasn't been trained on. These types of generalizations are the ultimate challenge for AI.

Newton: A Pre-Trained AI Foundation Model for Physical Behaviors

To address this challenge, Archetype AI pre-trained its model, Newton, using 0.59 billion samples from open-source datasets that cover a wide range of physical behaviors, including electrical currents, fluid flows, and sensor measurements. Newton uses a transformer-based deep neural network to make sense of this raw, noisy sensor data and uncover underlying patterns across a broad spectrum of physical phenomena.

Several lightweight, application-specific decoders were then trained to use the insights generated by Newton’s encoder. These decoders can predict future outcomes based on sensor data or reconstruct past events, applying the AI's findings to real-world tasks. Newton excels at real-time data analysis, whether it's from live sensors or prerecorded measurements.

Zero-Shot Forecasting in Physical AI

To test Newton’s capabilities, the team conducted simple physics experiments involving mechanical oscillations and thermodynamics—concepts familiar to most people from school. Even though Newton had never encountered these specific experiments, it was able to accurately predict the behavior of the systems in real-time, a technique known as zero-shot forecasting.

After success with these basic experiments, Newton was applied to more complex systems, such as predicting city electrical demand and daily temperature fluctuations. Newton's ability to generalize without additional training data allowed it to accurately forecast behaviors that would be challenging even for humans.

Generalization Beyond Targeted Training

What’s particularly exciting is that Newton consistently outperformed using zero-shot forecasting even when trained on specific datasets. For example, Newton, trained on a wide range of sensor data from all over the world, was more accurate in predicting the temperature of oil in electrical transformers than when it was trained specifically on transformer oil data. This suggests that physical AI foundation models like Newton possess powerful generalization capabilities, enabling them to understand systems far beyond their original training scope.

The Practical Implications of Physical AI

The success of models like Newton offers significant practical benefits:

• One Model, Many Use Cases: Instead of training separate AI models for each specific application, Newton can be applied across various domains, accelerating and simplifying the deployment of AI solutions for real-world problems.

• Reduced Training Data Requirements: A pre-trained physical AI model can accurately analyze sensor data with minimal additional training, making it valuable in situations where large datasets are difficult to obtain. This could pave the way for sensor-agnostic physical AI platforms capable of working with specialized sensors immediately, without the need for extensive retraining.

• Efficient Computing: The proposed architecture—combining a foundational encoder with lightweight, application-specific decoders—allows the model to be quickly adapted to various applications without significant changes to its core structure. This approach streamlines the creation of new use cases, reducing both computational demands and costs.

• Autonomous Learning from Observations: Newton’s ability to learn from raw observational data opens the door to autonomous systems that can adapt to new environments without human intervention, such as manual data labeling or model fine-tuning.

Unlocking New Possibilities with Physical AI

The potential applications of physical AI are vast, from building automotive monitoring systems and self-adapting robots to driving scientific discovery by uncovering new physical laws. If AI can discern the laws of nature through observation alone, what other hidden patterns could it reveal about the universe?