# AI-Newton: A Concept-Driven Physical Law Discovery System without Prior Physical Knowledge
**Authors**: You-Le Fang, Dong-Shan Jian, Xiang Li, Yan-Qing Ma
> eden@stu.pku.edu.cnSchool of Physics, Peking University, Beijing 100871, China
> dsjian@stu.pku.edu.cnSchool of Physics, Peking University, Beijing 100871, China
> lix-PHY@pku.edu.cnSchool of Physics, Peking University, Beijing 100871, China
> yqma@pku.edu.cnSchool of Physics, Peking University, Beijing 100871, ChinaCenter for High Energy Physics, Peking University, Beijing 100871, China
(December 11, 2025)
Abstract
While current AI-driven methods excel at deriving empirical models from individual experiments, a significant challenge remains in uncovering the common fundamental physics that underlie these modelsâa task at which human physicists are adept. To bridge this gap, we introduce AI-Newton, a novel framework for concept-driven scientific discovery. Our system autonomously derives general physical laws directly from raw, multi-experiment data, operating without supervision or prior physical knowledge. Its core innovations are twofold: (1) proposing interpretable physical concepts to construct laws, and (2) progressively generalizing these laws to broader domains. Applied to a large, noisy dataset of mechanics experiments, AI-Newton successfully rediscovers foundational and universal laws, such as Newtonâs second law, the conservation of energy, and the universal gravitation. This work represents a significant advance toward autonomous, human-like scientific discovery.
Introduction. â For centuries, the ultimate goal of fundamental physics research has been to describe a wide range of phenomena through a small number of discovered laws. Advances in artificial intelligence (AI) have made AI-driven scientific discovery a highly promising new paradigm [1]. Although AI has achieved remarkable results in tackling domain-specific challenges [2, 3], the ultimate aspiration from a paradigm-shifting perspective still lies in developing reliable AI systems capable of autonomous scientific discovery directly from a large collection of raw data without supervision [4, 5].
Current approaches to automated physics discovery focus on individual experiments, employing either neural network (NN)-based methods [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25] or symbolic techniques [26, 27, 28, 29, 30, 31, 32, 33]. By analyzing data from a single experiment, these methods can construct a specific model capable of predicting future data from the same experiment; if sufficiently simple, such a model may even be expressed in symbolic form [34, 35, 36]. Although these methods represent a crucial and successful stage towards automated scientific discovery, they have not yet reached a discovery capacity comparable to that of human physicists.
Human scientists advance further by discerning common patterns across specific models from different experiments and, on that basis, formulating general models that account for data from all such experiments. For instance, Newtonian mechanics provides a unifying and interpretable framework by defining meaningful physical concepts and formulating general laws that are valid across diverse phenomena. Therefore, a central challenge for the AI-driven physics discovery field is to evolve beyond problem-specific model fitting towards AI systems capable of discovering knowledge that is inherently generalizable and universally applicable.
In this Letter, we present AI-Newton, a concept-driven discovery system, which is designed for the critical question: how to extract concepts and general laws from problem-specific models. AI-Newton integrates an autonomous discovery workflow which is fundamentally built upon plausible reasoning and physical concepts. Given a collection of physical experiments, AI-Newton can gradually formulate a set of general laws applicable across a wide problem scope with neither supervision nor any prior physical knowledge. As a proof-of-concept implementation Code available at https://github.com/Science-Discovery/AI-Newton, by applying it to 46 different classical mechanics experiments, it can rediscover Newtonâs second law, energy conservation, law of gravitation and others in classical mechanics.
<details>
<summary>overview.png Details</summary>

### Visual Description
## Workflow Diagram: Autonomous Discovery Workflow
### Overview
The image is a workflow diagram illustrating an autonomous discovery process. It consists of three main sections: an "Experiment base" on the left, an "Autonomous discovery workflow" in the center, and a "Theory base" on the right. The diagram shows the flow of information and processes between these sections, highlighting the iterative nature of the discovery process.
### Components/Axes
* **Titles:**
* Experiment base
* Autonomous discovery workflow
* Theory base
* **Experiment base (Left):**
* Contains a list of experiments: Experiment 1, Experiment 2, ..., Experiment N.
* Experiment 1 contains: Physical objects, Geometric information, Experimental parameters, Space-time coordinates, Data generator.
* An arrow labeled "Experiments" points from the Experiment base to the Autonomous discovery workflow.
* **Autonomous discovery workflow (Center):**
* Consists of a series of steps:
* Selection: One experiment, A few concepts. Bounded by a dashed orange line.
* Search of physical laws: Extension of general laws, Direct search of specific laws. Bounded by a dashed gray line.
* Simplification and classification. Bounded by a dashed red line.
* Extraction of concepts and general laws. Bounded by a dashed blue line.
* Arrows indicate the flow of information between these steps.
* **Theory base (Right):**
* Consists of three components:
* Symbols
* Concepts: Dynamical concepts, Intrinsic concepts, Universal constants.
* Laws: Specific laws, General laws.
* Arrows labeled "represent" and "extract" show the relationship between Symbols, Concepts, and Laws.
* Arrows labeled "Concepts" and "Laws" point from the Theory base to the Autonomous discovery workflow.
* **Legend (Bottom):**
* Recommendation engine (dashed orange line)
* Symbolic regression (dashed gray line)
* Differential algebra & variable control (dashed red line)
* Plausible reasoning (dashed blue line)
### Detailed Analysis or ### Content Details
* **Experiment Base:**
* The Experiment base provides the initial data and parameters for the discovery process.
* Experiment 1 lists specific data types: Physical objects, Geometric information, Experimental parameters, Space-time coordinates, Data generator.
* The presence of "Experiment 2" and "Experiment N" suggests that multiple experiments can be used.
* **Autonomous Discovery Workflow:**
* The workflow starts with "Selection," where one experiment and a few concepts are chosen.
* The next step is "Search of physical laws," which involves extending general laws and directly searching for specific laws.
* "Simplification and classification" follows, likely to reduce complexity and categorize the findings.
* The final step is "Extraction of concepts and general laws," where the discovered concepts and laws are formalized.
* **Theory Base:**
* The Theory base represents the existing knowledge and framework used in the discovery process.
* Symbols, Concepts, and Laws are interconnected, with "represent" and "extract" arrows indicating the flow of information between them.
* Concepts are further divided into Dynamical concepts, Intrinsic concepts, and Universal constants.
* Laws are divided into Specific laws and General laws.
* **Flow of Information:**
* Experiments from the Experiment base feed into the Autonomous discovery workflow.
* Concepts and Laws from the Theory base also feed into the Autonomous discovery workflow.
* The workflow is iterative, with information flowing between the steps.
### Key Observations
* The diagram emphasizes the iterative nature of the autonomous discovery process.
* The process involves both data from experiments and existing theoretical knowledge.
* The workflow is structured into distinct steps, each with a specific purpose.
* The legend provides context for the different components of the workflow.
### Interpretation
The diagram illustrates a systematic approach to autonomous scientific discovery. It combines experimental data with theoretical frameworks to identify and formalize new concepts and laws. The iterative nature of the workflow allows for continuous refinement and improvement of the discovered knowledge. The diagram highlights the importance of both data-driven and theory-driven approaches in scientific discovery. The use of different colored dashed lines to represent different engines/methods suggests that each step in the workflow can be approached using different computational techniques.
</details>
Figure 1: AI-Newtonâs experiment base, theory base, and autonomous discovery workflow.
Knowledge base and knowledge representation. â AI-Newton contains an experiment base and a theory base, as shown in Fig. 1. The experiment base stores physical experiments and corresponding simulated data generators. The inputs for each experiment include only the physical objects involved, geometric information, experimental parameters, and space-time coordinates, which define an experiment. To emphasize that no prior physical knowledge is used, all other concepts, such as mass or energy, are autonomously discovered in AI-Newton. The output of each experiment is simulated data with statistical errors.
The theory base stores physical knowledge explicitly in an interconnected library of symbols, concepts, and laws. This design mirrors how human physicists construct concise, universal laws from conceptual building blocks. In contrast to prior work, which interprets latent features in NNs as physical concepts [37, 23, 38], AI-Newton represents concepts and laws in an explicit, symbolic form. This greatly enhances interpretability and makes the acquired knowledge easier to transfer to new problems. Moreover, the introduction of powerful intermediate concepts allows complex physical laws to be expressed concisely, which in turn makes them more amenable to discovery through techniques like symbolic regression (SR) [26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36]. Initially, the concept layer contains only space-time coordinates; new concepts are autonomously defined and registered using a dedicated physical domain-specific language (DSL). (See Supplemental Materials (SMs) [39] for details.)
A robust knowledge representation is crucial because our goal is for the AI to discover generalizable knowledge across diverse systems, which requires transferring knowledge between different problems. To achieve this, we designed a physical DSL with a well-defined structure. This DSL not only formulates equations but also encodes the properties of physical objects and the relationships between physical quantities. For instance, given the known concepts of coordinate $x$ and time $t$ , the velocity of a ball can be defined in the DSL as:
$$
C_{1}:=\forall i\text{: Ball},\,\mathrm{d}x[i]/\mathrm{d}t, \tag{1}
$$
where $i$ indexes the balls and $C_{1}$ denotes the symbol of velocity, with the subscript $1$ varying across tests. In addition to dynamical concepts like velocity, the system also automatically identifies two other types: intrinsic concepts (e.g., mass, spring constant), which depend solely on specific physical objects, and universal constants (e.g., the gravitational constant), which are independent of all other quantities. Both are defined by documenting their measurement procedures. For example, mass of a ball could be defined as:
$$
\begin{split}C_{2}:=&\forall i\text{: Ball},\text{Intrinsic}[\\
&\text{ExpName}(o_{1}\rightarrow i,o_{2}\to s),\text{L}[s]-\text{L}_{0}[s]],\end{split} \tag{2}
$$
where ExpName is the name of an experiment. In this experiment, the measured ball $i$ is suspended from a fixed spring $s$ , and the spring elongation $\text{L}[s]-\text{L}_{0}[s]$ serves as the measurement of the mass. Recording the measurement procedures of intrinsic concepts is essential, since it allows the value of an intrinsic property to be retrieved by invoking its defining experiment, ensuring conceptual consistency across different problems.
These explicit concepts serve as the building blocks for the laws layer, which stores discovered physical laws, such as conserved quantities and dynamical equations. The laws are categorized into specific laws (valid for one experiment with specific forms) and general laws (valid across diverse experiments with general forms). Within this framework, prior research in AI-driven physics discovery has concentrated on identifying specific laws. The introduction of general laws enables AI-Newton to simultaneously describe physics in various complex systems with compact and concise formulations. For instance, consider a system with a ball on an inclined plane connected to a fixed end via a spring. By applying the general law discovered by AI-Newton (Newtonâs second law in the $x$ -direction):
$$
\forall i:\text{Ball},\,m_{i}a_{i,x}+(\nabla_{i}V_{g})_{x}+(\nabla_{i}V_{k})_{x}=0, \tag{3}
$$
the more complex dynamical equation of the ball can be concretely derived as:
$$
\begin{split}&ma_{x}-\frac{c_{x}c_{z}}{c_{x}^{2}+c_{y}^{2}+c_{z}^{2}}mg\\
+&\frac{\left[\left(c_{y}^{2}+c_{z}^{2}\right)x-c_{x}\left(c_{y}y+c_{z}z\right)\right]}{\left(c_{x}^{2}+c_{y}^{2}+c_{z}^{2}\right)L}k\Delta{L}=0,\end{split} \tag{4}
$$
where $(c_{x},c_{y},c_{z})$ is the normal vector defining the inclined plane. For multi-object systems, concrete dynamical equations can be much more complex than the general laws, making them hard to be obtained using previous symbolic approaches. These cases highlight the efficacy of our concept-driven hierarchical approach.
Autonomous discovery workflow. â The autonomous discovery workflow in AI-Newton continuously distill knowledgeâexpressed as physical concepts and lawsâfrom experimental data, as shown in Fig. 1. Plausible reasoning, a method based on rational inference from partial evidence [40, 41], is the key to discovering knowledge. Unlike deductive logic, it produces contextually reasonable rather than universally certain conclusions, mirroring scientific practice where hypotheses precede rigorous verification.
The workflow initiates each trial by selecting an experiment and a few concepts from the theory base. This selection is governed by a recommendation engine that integrates a UCB-inspired value function [42, 43, 44, 45, 46] with a dynamically adapted NN. The NNâs architecture is updated in real-time to favor configurations that lead to efficient knowledge extraction. This mechanism enables the system to emulate human-like learning, naturally balancing the trade-off between exploration and exploitation.
To ensure the workflow establishes foundational knowledge before tackling complex experiments, we introduce an era-control strategy. Within a given era, every trial must conclude within a specific wall-clock time limit. If no new knowledge is acquired after a sufficient number of trials, the system advances to a new era with an exponentially increased time limit. Consequently, this strategy keeps the system focused on simpler experiments in the early phases. (See SMs [39] for more details.)
The next step of each trial is to explore new laws from the selected experiment and concepts. Specific laws can be discovered through direct searching for relations among the selected concepts within the allowed operational space, which is nothing but SR. Our SR implementation combines direct instantiation-verification and PCA-based differential polynomial regression [47, 48, 49, 50]. Furthermore, new general laws may emerge by extending existing ones through plausible reasoning. The core idea of plausible reasoning here is that, if a general law holds across multiple experiments but fails in the current one, there is a possibility to derive a valid modified law by adding simple terms to the original formulation via SR. For instance, while kinetic energy conservation governs elastic collisions, it fails in spring systems. Through plausible reasoning, AI-Newton introduces additional terms (elastic potential) to restore conservation. Mirroring human research practice, the system heuristically leverages existing general laws and selected concepts to search for physical laws that explain new experimental data.
The aforementioned process may generate redundant knowledge causing an explosion in both the theory base and search space that severely hinders continuous discovery under limited resources. To address this, AI-Newton simplifies physical laws into minimal representations in each trial. For the example shown in this paper, we employ the Rosenfeld Gröbner algorithm [51, 52, 53, 54] from differential algebra to perform the simplification (See SMs [39] for more details). Furthermore, through controlled-variable analysis, AI-Newton numerically identifies the dependencies of relations on physical objects and experimental parameters, using these dependencies as the basis for classification.
After identifying new laws, AI-Newton extracts new concepts from the processed results through plausible reasoning: a conserved quantity in the current experiment suggests broader utility, triggering its extraction as a new concept. Similarly, it proposes new general laws from directly-searched specific laws that also hold in multiple other experiments. All accumulated knowledge are updated to the theory base.
<details>
<summary>test_cases.png Details</summary>

### Visual Description
## Physical Laws and Experiments
### Overview
The image presents a visual and mathematical overview of physical laws, linking physical objects to schematic experiments and the corresponding mathematical formulations. It is divided into three main sections: "Physical objects," "Schematic of experiments," and "Discovered important general laws."
### Components/Axes
* **Title:** Physical objects, Schematic of experiments, Discovered important general laws
* **Physical Objects:** Shows a sphere, a spring, and a triangular prism.
* **Schematic of Experiments:** Contains 8 diagrams illustrating different physical setups. These are numbered (1) through (8), with (9) being separate.
* **Discovered Important General Laws:** Presents mathematical equations related to energy conservation, universal gravitation, and Newton's second law.
### Detailed Analysis or ### Content Details
**Physical Objects:**
* A white sphere.
* A metallic spring.
* A gray triangular prism.
**Schematic of Experiments:**
The "Schematic of experiments" section is divided into a 3x3 grid, with each cell containing a diagram.
* **(1):** A sphere moving towards a spring.
* **(2):** Three spheres moving towards each other, connected by springs.
* **(3):** Two rows of spheres connected by springs, with arrows indicating oscillation.
* **(4):** Three spheres connected in a triangle by springs, with arrows indicating oscillation and rotation.
* **(5):** A sphere falling under gravity near Earth's surface.
* **(6):** A sphere rolling down an inclined plane.
* **(7):** A series of spheres connected by springs, oscillating near a surface.
* **(8):** A sphere attached to a spring, rolling down an inclined plane.
* **(9):** Four spheres connected by dotted lines, with arrows indicating rotation, representing universal gravitation.
**Discovered Important General Laws:**
* **Energy Conservation:**
* Equation: â(T\_k) + â(ÎŽ\_λ V\_λ) = const., where k â {x, y, z} and λ â {k, g, G}
* "where T\_k and V\_λ are defined as:"
* **Kinetic Energy (T\_k):**
* Equation: T\_k = â(m\_i v\_{i,k}^2), where i â Particles
* **Potential Energy (Springs) (V\_k):**
* Equation: V\_k = â(k\_i (L\_i - L\_{0,i})^2), where i â Springs
* **Potential Energy (Gravity) (V\_g):**
* Equation: V\_g = â(2 m\_i g z\_i), where i â Particles
* **Potential Energy (Universal Gravitation) (V\_G):**
* Equation: V\_G = â(2 (-G m\_i m\_j / r\_{ij})), where i, j â Particles
* **Newton's Second Law:**
* Equation: 2a\_k + â(ÎŽ\_λ (1/m âV\_λ / âÎș)) = 0, where λ â {k, g, G} and Îș â {x, y, z}
* "(Ύ\_λ = 0 or 1, determined spontaneously during instantiation as specific laws in experiments)"
### Key Observations
* The image connects physical objects to simplified experimental setups and their corresponding mathematical descriptions.
* The schematic experiments illustrate basic physics concepts such as motion, oscillation, gravity, and inclined planes.
* The mathematical equations represent fundamental laws of physics, including energy conservation, universal gravitation, and Newton's second law.
### Interpretation
The image serves as a visual and mathematical summary of fundamental physics principles. It demonstrates how real-world objects and experimental setups can be modeled mathematically using established physical laws. The progression from physical objects to schematic experiments to mathematical equations highlights the process of abstraction and formalization in physics. The equations provide a quantitative framework for understanding and predicting the behavior of the systems depicted in the schematic experiments. The image is a concise representation of the relationship between physical phenomena, experimental observation, and mathematical theory.
</details>
Figure 2: Schematic of tested experiments and main general laws discovered. Some complex configurations are omitted for clarity. See text for details.
Rediscovering Laws of Newtonian Mechanics. â To evaluate AI-Newtonâs performance, we apply it to Newtonian mechanics problems, focusing on a set of 46 predefined experiments. These experiments involve three primary types of physical objects: balls (either small balls or celestial bodies), springs, and inclined planes. The experiments are designed to investigate both isolated and coupled systems, as illustrated in Fig. 2, including:
1. Free motion of individual balls and springs;
1. Elastic collision of balls;
1. Coupled systems demonstrating translational vibrations, rotational oscillations, and pendulum-like motions;
1. Gravity-related problems, such as projectile motion and motion on inclined planes, along with complex spring-ball systems;
1. Celestial mechanics problems involving gravitational interactions.
The complexities of experiments are systematically increased by varying the number of physical objects and spatial dimensions, encompassing high-degree-of-freedom problems such as coupled oscillations of chained 2-ball-2-spring systems on inclined planes, rotational dynamics of 4-ball-4-spring systems, and other complex configurations. To simulate realistic experimental conditions, all test data are generated by solving differential equations and incorporating Gaussian-distributed errors. This comprehensive experimental setup covers three types of forces in Newtonian mechanics, elastic forces, gravity near Earthâs surface, and universal gravitational forces, while incorporating realistic measurement uncertainties. In this way, it enables rigorous evaluation of AI-Newtonâs capability to discover physical laws from noisy experimental data.
We evaluated the performance of our proof-of-concept implementation on an Intel Xeon Platinum 8370C (128 threads @ 3.500GHz) platform with NVIDIA A40 GPU, configured with 64 cores for parallel processing. With max trials set to 1200 and an average runtime of 48 hours, the system demonstrated robust knowledge discovery capabilities, identifying approximately 90 physical concepts and 50 general laws on average across the test cases. The discoveries include significant general laws such as energy conservation and Newtonâs second law along with their relevant concepts, as shown in Fig. 2, providing complete explanatory for all experiments covering systems from simple to high-degree-of-freedom complex configurations.
<details>
<summary>knowledge_progression.png Details</summary>

### Visual Description
## Dot Plot: Number of Trials for Concept Learning
### Overview
The image is a dot plot showing the number of trials required to learn different concepts. The x-axis represents the number of trials, and the y-axis lists the concepts. Each concept has a dot representing the mean number of trials, with error bars indicating the variability. The plot is divided into six regions, labeled I through VI, with varying shades of green.
### Components/Axes
* **X-axis:** "Number of trials", ranging from 0 to 800, with tick marks at intervals of 100.
* **Y-axis:** "Concepts", listing the following concepts from bottom to top: m, V, k, g, a, T, P, Vk, Vg, Fk, Fg, G, VG, FG.
* **Data Points:** Black dots represent the mean number of trials for each concept.
* **Error Bars:** Gray horizontal lines extending from each dot, indicating the range of trials.
* **Background Regions:** Six vertical regions labeled I, II, III, IV, V, and VI, shaded with different intensities of green. The Roman numerals are in light gray.
### Detailed Analysis
Here's a breakdown of the data for each concept, including the approximate mean number of trials and the range indicated by the error bars:
* **m:** Mean ~60 trials, Range ~20-100 trials.
* **V:** Mean ~70 trials, Range ~30-110 trials.
* **k:** Mean ~80 trials, Range ~30-130 trials.
* **g:** Mean ~220 trials, Range ~150-290 trials.
* **a:** Mean ~240 trials, Range ~200-280 trials.
* **T:** Mean ~310 trials, Range ~280-340 trials.
* **P:** Mean ~350 trials, Range ~300-400 trials.
* **Vk:** Mean ~400 trials, Range ~350-450 trials.
* **Vg:** Mean ~420 trials, Range ~380-460 trials.
* **Fk:** Mean ~450 trials, Range ~400-500 trials.
* **Fg:** Mean ~500 trials, Range ~450-550 trials.
* **G, VG:** Mean ~750 trials, Range ~700-800 trials.
* **FG:** Mean ~780 trials, Range ~730-830 trials.
The background regions are defined as follows:
* **Region I:** 0-150 trials
* **Region II:** 150-225 trials
* **Region III:** 225-300 trials
* **Region IV:** 300-400 trials
* **Region V:** 400-650 trials
* **Region VI:** 650-850 trials
### Key Observations
* The number of trials required to learn the concepts varies significantly.
* Concepts 'm', 'V', and 'k' require the fewest trials, while 'G, VG' and 'FG' require the most.
* The error bars indicate variability in the number of trials needed for each concept.
* The concepts are generally ordered from easiest to learn (fewest trials) to hardest to learn (most trials).
### Interpretation
The dot plot visualizes the difficulty of learning different concepts, as measured by the number of trials required. The data suggests a clear hierarchy in concept difficulty, with some concepts being learned much faster than others. The error bars provide insight into the consistency of learning for each concept. The background regions (I-VI) could represent different phases or levels of learning difficulty, with each concept falling into a specific region based on the number of trials needed for mastery. The plot could be used to compare the relative difficulty of different concepts and to identify concepts that are particularly easy or difficult to learn.
</details>
Figure 3: Statistical analysis of concept discovery timing on 10 test cases, recording the mean and standard deviation of discovery timings for key concepts. Number of trials means the number of analysis trial attempt has been done, not distinguishing which experiment. Roman numerals (I, II, âŠ) in the background indicate the eras defined by the era-control strategy.
Statistical discovery progression on 10 test cases is illustrated in Fig. 3, showing the timing distribution of important concept discoveries. This discovery progression exhibits an incremental pattern, where AI-Newton first explores simple concepts (e.g., mass) before advancing to more complex ones (e.g., force). For instance, gravitational acceleration $g$ is defined as a constant by analyzing free-fall or projectile motion, where the vertical acceleration $a_{z}$ of the ball is invariant. In experiments with elastic collisions between balls, conservation of kinetic energy $T$ is discovered and proposed as a general law. Through plausible reasoning, elastic potential energy $V_{k}$ , gravitational potential energy near Earthâs surface $V_{g}$ , and universal gravitational potential energy $V_{G}$ are progressively defined when trying to apply the conservation of kinetic energy to inelastic experiments. These are then incorporated with kinetic energy conservation to ultimately formulate the complete law of energy conservation. The discovery of Newtonâs second law follows an analogous progression: it is first proposed in a simple experimental context and then generalized through plausible reasoning.
It is important to emphasize that the system is able to independently discover and unify fundamental concepts from disparate physical contexts. For instance, AI-Newton can derive the concept of âmassâ through two distinct experimental routes: from the static elongation of a spring under gravity (defining gravitational mass, $m_{g}$ ) and from the experiment of a horizontal spring-mass oscillation system (defining inertial mass, $m_{i}$ ). Critically, the system then autonomously verify the numerical equivalence of $m_{g}$ and $m_{i}$ , effectively indicating a cornerstone of general relativityâthe weak equivalence principleâfrom raw data alone.
Summary. â We introduce AI-Newton, a novel framework for the autonomous discovery of general physical laws from raw data across a large set of experiments, without supervision or pre-existing physical knowledge. This approach transcends current AI-driven methods, which are limited to extracting specific laws from individual experiments. Our main contributions are based on plausible reasoning, enabling us to: (1) propose physical concepts from the extracted laws; and (2) extend an existing general law by adding new terms, thereby adapting it to describe a wider range of experiments. Introducing interpretable physical concepts allows discovered laws to remain concise, making them more tractable for SR to identify. Furthermore, iteratively constructing general laws from existing ones enables a gradual, scalable discovery process. Applied to a large, noisy dataset of mechanics experiments, AI-Newton successfully rediscovers foundational laws, including Newtonâs second law, the conservation of energy, and the law of universal gravitation. This work thus offers a promising pathway toward building AI systems capable of contributing to frontier scientific research.
As a first step, we employ AI-Newton to rediscover known physical lawsâa task where direct reliance on large language models (LLMs) is unsuitable, as they already possess this knowledge. In future applications to frontier science, however, the DSL, the recommendation engine and the plausible reasoning components of the framework could be replaced or augmented by LLMs. This integration would grant the system direct access to all existing knowledge, enabling a more informed and efficient discovery process.
Acknowledgements. We would like to thank Hong-Fei Zhang for early participant of the project and many valuable discussions. This work is supported by the National Natural Science Foundation of China (No. 12325503), and the High-performance Computing Platform of Peking University.
References
- [1] Y. Xu, X. Liu, X. Cao, C. Huang, E. Liu, S. Qian, X. Liu, Y. Wu, F. Dong, C.-W. Qiu, et al., Artificial intelligence: A powerful paradigm for scientific research, The Innovation 2 (2021) .
- [2] H. Wang, T. Fu, Y. Du, W. Gao, K. Huang, Z. Liu, P. Chandak, S. Liu, P. Van Katwyk, A. Deac, et al., Scientific discovery in the age of artificial intelligence, Nature 620 (2023) 47â60.
- [3] X. Zhang, L. Wang, J. Helwig, Y. Luo, C. Fu, Y. Xie, M. Liu, Y. Lin, Z. Xu, K. Yan, et al., Artificial intelligence for science in quantum, atomistic, and continuum systems, Foundations and TrendsÂź in Machine Learning 18 (2025) 385â912.
- [4] C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha, The ai scientist: Towards fully automated open-ended scientific discovery, [arXiv:2408.06292].
- [5] C. K. Reddy and P. Shojaee, Towards scientific discovery with generative ai: Progress, opportunities, and challenges, Proceedings of the AAAI Conference on Artificial Intelligence 39 (Apr., 2025) 28601â28609.
- [6] M. Schmidt and H. Lipson, Distilling free-form natural laws from experimental data, science 324 (2009) 81â85.
- [7] S. Brunton, J. Proctor, and J. Kutz, Discovering governing equations from data: Sparse identification of nonlinear dynamical systems, Proceedings of the National Academy of Sciences 113 (09, 2015) 3932â3937.
- [8] K. Champion, B. Lusch, J. N. Kutz, and S. L. Brunton, Data-driven discovery of coordinates and governing equations, Proceedings of the National Academy of Sciences 116 (2019) 22445â22451.
- [9] T. Wu and M. Tegmark, Toward an artificial intelligence physicist for unsupervised learning, Physical Review E 100 (2019) 033311.
- [10] S. Greydanus, M. Dzamba, and J. Yosinski, Hamiltonian neural networks. Curran Associates Inc., Red Hook, NY, USA, 2019.
- [11] M. Cranmer, S. Greydanus, S. Hoyer, P. Battaglia, D. Spergel, and S. Ho, Lagrangian neural networks, [arXiv:2003.04630].
- [12] B. M. De Silva, D. M. Higdon, S. L. Brunton, and J. N. Kutz, Discovery of physics from data: Universal laws and discrepancies, Frontiers in artificial intelligence 3 (2020) 25.
- [13] Z. Liu and M. Tegmark, Machine learning conservation laws from trajectories, Physical Review Letters 126 (2021) 180604.
- [14] G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, Physics-informed machine learning, Nature Reviews Physics 3 (2021) 422â440.
- [15] Z. Liu, V. Madhavan, and M. Tegmark, Machine learning conservation laws from differential equations, Physical Review E 106 (2022) 045307.
- [16] G. Camps-Valls, A. Gerhardus, U. Ninad, G. Varando, G. Martius, E. Balaguer-Ballester, R. Vinuesa, E. Diaz, L. Zanna, and J. Runge, Discovering causal relations and equations from data, Physics Reports 1044 (2023) 1â68.
- [17] C. Cornelio, S. Dash, V. Austel, T. R. Josephson, J. Goncalves, K. L. Clarkson, N. Megiddo, B. El Khadir, and L. Horesh, Combining data and theory for derivable scientific discovery with ai-descartes, Nature Communications 14 (2023) 1777.
- [18] P. Lemos, N. Jeffrey, M. Cranmer, S. Ho, and P. Battaglia, Rediscovering orbital mechanics with machine learning, Machine Learning: Science and Technology 4 (2023) 045002.
- [19] Z. Liu, P. O. Sturm, S. Bharadwaj, S. J. Silva, and M. Tegmark, Interpretable conservation laws as sparse invariants, Phys. Rev. E 109 (Feb, 2024) L023301.
- [20] R. Cory-Wright, C. Cornelio, S. Dash, B. El Khadir, and L. Horesh, Evolving scientific discovery by unifying data and background knowledge with ai hilbert, Nature Communications 15 (2024) 5922.
- [21] D. Zheng, V. Luo, J. Wu, and J. B. Tenenbaum, Unsupervised learning of latent physical properties using perception-prediction networks, [arXiv:1807.09244].
- [22] M. Tegmark, Latent Representations of Dynamical Systems: When Two is Better Than One, [arXiv:1902.03364].
- [23] R. Iten, T. Metger, H. Wilming, L. Del Rio, and R. Renner, Discovering physical concepts with neural networks, Physical review letters 124 (2020) 010508.
- [24] B. Chen, K. Huang, S. Raghupathi, I. Chandratreya, Q. Du, and H. Lipson, Automated discovery of fundamental variables hidden in experimental data, Nature Computational Science 2 (2022) 433â442.
- [25] Q. Li, T. Wang, V. Roychowdhury, and M. K. Jawed, Metalearning generalizable dynamics from trajectories, Physical Review Letters 131 (2023) 067301.
- [26] S.-M. Udrescu and M. Tegmark, Ai feynman: A physics-inspired method for symbolic regression, Science Advances 6 (2020) eaay2631.
- [27] S.-M. Udrescu, A. Tan, J. Feng, O. Neto, T. Wu, and M. Tegmark, Ai feynman 2.0: Pareto-optimal symbolic regression exploiting graph modularity, Advances in Neural Information Processing Systems 33 (2020) 4860â4871.
- [28] T. Bendinelli, L. Biggio, and P.-A. Kamienny, Controllable neural symbolic regression, in International Conference on Machine Learning, pp. 2063â2077, PMLR. 2023.
- [29] W. Tenachi, R. Ibata, and F. I. Diakogiannis, Deep symbolic regression for physics guided by units constraints: toward the automated discovery of physical laws, The Astrophysical Journal 959 (2023) 99.
- [30] Y. Tian, W. Zhou, M. Viscione, H. Dong, D. Kammer, and O. Fink, Interactive symbolic regression with co-design mechanism through offline reinforcement learning, Nature Communications 16 (04, 2025) .
- [31] M. Cranmer, Interpretable machine learning for science with pysr and symbolicregression. jl, [arXiv:2305.01582].
- [32] M. Du, Y. Chen, Z. Wang, L. Nie, and D. Zhang, Large language models for automatic equation discovery of nonlinear dynamics, Physics of Fluids 36 (2024) .
- [33] B. Romera-Paredes, M. Barekatain, A. Novikov, M. Balog, M. P. Kumar, E. Dupont, F. J. Ruiz, J. S. Ellenberg, P. Wang, O. Fawzi, et al., Mathematical discoveries from program search with large language models, Nature 625 (2024) 468â475.
- [34] M. Valipour, B. You, M. Panju, and A. Ghodsi, Symbolicgpt: A generative transformer model for symbolic regression, arXiv:2106.14131 (2021) .
- [35] X. Chu, H. Zhao, E. Xu, H. Qi, M. Chen, and H. Shao, Neural symbolic regression using control variables, [arXiv:2306.04718].
- [36] S. MeĆŸnar, S. DĆŸeroski, and L. Todorovski, Efficient generator of mathematical expressions for symbolic regression, Machine Learning 112 (2023) 4563â4596.
- [37] C. Wang, H. Zhai, and Y.-Z. You, Emergent schrödinger equation in an introspective machine learning architecture, Science Bulletin 64 (2019) 1228â1233.
- [38] B.-B. Li, Y. Gu, and S.-F. Wu, Discover physical concepts and equations with machine learning, [arXiv:2412.12161].
- [39] âSee the arXiv supplemental material for details on the domain-specific language, recommendation engine, and computational engine.â.
- [40] G. PĂłlya, Mathematics and plausible reasoning: Induction and analogy in mathematics, vol. 1. Princeton University Press, 1990.
- [41] G. PĂłlya, Mathematics and Plausible Reasoning: Patterns of plausible inference, vol. 2. Princeton University Press, 1990.
- [42] T. Lai and H. Robbins, Asymptotically efficient adaptive allocation rules, Adv. Appl. Math. 6 (Mar., 1985) 4â22.
- [43] T. L. Lai, Adaptive treatment allocation and the multi-armed bandit problem, The Annals of Statistics 15 (1987) 1091â1114. http://www.jstor.org/stable/2241818.
- [44] R. Agrawal, Sample mean based index policies by o (log n) regret for the multi-armed bandit problem, Advances in applied probability 27 (1995) 1054â1078.
- [45] M. N. Katehakis and H. Robbins, Sequential choice from several populations., Proceedings of the National Academy of Sciences 92 (1995) 8584â8585.
- [46] P. Auer, Using confidence bounds for exploitation-exploration trade-offs, Journal of Machine Learning Research 3 (2002) 397â422. https://www.jmlr.org/papers/volume3/auer02a/auer02a.pdf.
- [47] K. Pearson, Liii. on lines and planes of closest fit to systems of points in space, The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science 2 (1901) 559â572.
- [48] L. Wang, Discovering phase transitions with unsupervised learning, Phys. Rev. B 94 (Nov, 2016) 195105.
- [49] H. Kiwata, Deriving the order parameters of a spin-glass model using principal component analysis, Physical Review E 99 (2019) 063304.
- [50] D. Yevick, Conservation laws and spin system modeling through principal component analysis, Computer Physics Communications 262 (2021) 107832.
- [51] F. Boulier, D. Lazard, F. Ollivier, and M. Petitot, Representation for the radical of a finitely generated differential ideal, in Proceedings of the 1995 international symposium on Symbolic and algebraic computation, ISSAC â95, pp. 158â166. Association for Computing Machinery, New York, NY, USA, 1995.
- [52] F. Boulier, D. Lazard, F. Ollivier, and M. Petitot, Computing representations for radicals of finitely generated differential ideals, Applicable Algebra in Engineering, Communication and Computing 20 (2009) 73â121.
- [53] Maplesoft, Differential algebra in maple, Maplesoft Help Center (2024) . https://cn.maplesoft.com/support/help/Maple/view.aspx?path=DifferentialAlgebra.
- [54] Maplesoft, Maple 2024, 2024. https://www.maplesoft.com/products/maple/.