# 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
> School of Physics, Peking University, Beijing 100871, China
> School of Physics, Peking University, Beijing 100871, China Center 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
## Flowchart Diagram: Autonomous Discovery Workflow System
### Overview
The diagram illustrates a three-stage system for autonomous scientific discovery, connecting experimental data processing (left), algorithmic workflow (center), and theoretical knowledge representation (right). The system uses color-coded arrows to represent different computational methods.
### Components/Axes
1. **Experiment Base (Green Section)**
- Contains three experiments (1, 2, N) with identical structure:
- Physical objects
- Geometric information
- Experimental parameters
- Space-time coordinates
- Data generator
- Arrows labeled "Experiments" point rightward to the workflow section
2. **Autonomous Discovery Workflow (Purple Section)**
- Contains four vertically stacked components:
- **Selection** (dashed orange border)
- Contains: "One experiment", "A few concepts"
- **Search of physical laws** (dashed gray border)
- Subcomponents:
- Extension of general laws
- Direct search of specific laws
- **Simplification and classification** (dashed red border)
- **Extraction of concepts and general laws** (dashed blue border)
- Arrows connect components vertically with different styles:
- Solid black (Selection â Search)
- Dashed gray (Search â Simplification)
- Dashed red (Simplification â Extraction)
3. **Theory Base (Blue Section)**
- Contains three hierarchical components:
- **Symbols** (top)
- Arrows: "represent" (down), "extract" (up)
- **Concepts** (middle)
- Subcategories:
- Dynamical concepts
- Intrinsic concepts
- Universal constants
- **Laws** (bottom)
- Subcategories:
- Specific laws
- General laws
- Arrows connect components vertically with "represent" (down) and "extract" (up) labels
4. **Legend (Bottom)**
- Four computational methods with color coding:
- Recommendation engine (yellow square)
- Symbolic regression (gray square)
- Differential algebra & variable control (red square)
- Plausible reasoning (blue square)
- Corresponds to arrow styles in workflow:
- Blue dashed arrows = Plausible reasoning
- Red dashed arrows = Differential algebra
- Gray dashed arrows = Symbolic regression
- Yellow solid arrows = Recommendation engine
### Detailed Analysis
- **Experiment Base**: Standardized experimental data format across all experiments
- **Workflow Flow**:
1. Experiments â Selection (orange dashed border)
2. Selection â Search (gray dashed)
3. Search â Simplification (red dashed)
4. Simplification â Extraction (blue dashed)
- **Theory Base Hierarchy**:
Symbols â Concepts â Laws (bottom-up)
Concepts/Laws â Symbols (top-down via "extract" arrows)
### Key Observations
1. Color-coded arrows maintain consistent methodology throughout:
- Blue dashed arrows (Plausible reasoning) connect Extraction to Theory base
- Red dashed arrows (Differential algebra) connect Search to Simplification
2. Bidirectional arrows between Concepts and Symbols indicate dynamic knowledge representation
3. Experimental data flows through all workflow stages before reaching theory
4. Theory base maintains both specific/general laws and dynamical/intrinsic concepts
### Interpretation
This system demonstrates a closed-loop scientific discovery process where:
1. Experimental data (green) is processed through algorithmic methods (colored arrows)
2. Workflow stages progressively abstract from specific experiments to general laws
3. Theoretical knowledge (blue) feeds back into experimental design through concept extraction
4. The color-coded methodology suggests:
- Plausible reasoning (blue) handles high-level abstraction
- Differential algebra (red) manages mathematical formalism
- Symbolic regression (gray) deals with pattern recognition
- Recommendation engine (yellow) optimizes experimental selection
The bidirectional arrows between Concepts and Symbols suggest an iterative knowledge refinement process, while the vertical workflow progression indicates increasing abstraction from raw data to universal laws.
</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
## Diagram: Physical Systems and Derived Physical Laws
### Overview
The image presents a technical diagram illustrating physical systems, experimental schematics, and derived physical laws. It is divided into three vertical sections:
1. **Physical objects** (left)
2. **Schematic of experiments** (center)
3. **Discovered important general laws** (right)
The diagram combines visual representations of physical systems with mathematical formulations of physical laws.
---
### Components/Axes
#### Physical Objects (Left Section)
1. **Sphere**: A simple 3D sphere labeled as a physical object.
2. **Spring**: A helical spring depicted in a compressed/extended state.
3. **Wedge**: A triangular prism representing an inclined plane.
#### Schematic of Experiments (Center Section)
Nine labeled diagrams (1â9) illustrate experimental setups:
- **(1)**: Sphere with a rightward arrow (linear motion).
- **(2)**: Spring with bidirectional arrows (compression/expansion).
- **(3)**: Stacked springs and particles with bidirectional arrows (interaction).
- **(4)**: Spring-chain system with bidirectional arrows (coupled motion).
- **(5)**: Sphere near Earth's surface with a downward arrow (gravity).
- **(6)**: Sphere on an inclined plane with a rightward arrow (gravity-induced motion).
- **(7)**: Sphere on an inclined plane with a spring (gravity + spring interaction).
- **(8)**: Sphere on an inclined plane with two springs (coupled forces).
- **(9)**: Four-particle system with bidirectional arrows (universal gravitation).
#### Discovered Laws (Right Section)
- **Energy Conservation**:
Equation:
$$
\sum_{\kappa \in \{x,y,z\}} T_\kappa + \sum_{\lambda \in \{k,g,G\}} \delta_\lambda V_\lambda = \text{const.}
$$
Definitions:
- $ T_\kappa = \sum_{i \in \text{Particles}} m_i v_{i,\kappa}^2 $ (kinetic energy)
- $ V_k = \sum_{i \in \text{Springs}} k_i (L_i - L_{0,i})^2 $ (spring potential energy)
- $ V_g = \sum_{i \in \text{Particles}} 2m_i g z_i $ (gravitational potential energy)
- $ V_G = \sum_{i,j \in \text{Particles}} 2\left(-\frac{G m_i m_j}{r_{ij}}\right) $ (universal gravitation potential energy)
- **Newton's Second Law**:
Equation:
$$
2a_\kappa + \sum_{\lambda \in \{k,g,G\}} \delta_\lambda \left(\frac{1}{m} \frac{\partial V_\lambda}{\partial \kappa}\right) = 0, \quad \kappa \in \{x,y,z\}
$$
Note: $ \delta_\lambda = 0 $ or $ 1 $, determined during experimentation.
---
### Detailed Analysis
#### Physical Objects
- The sphere, spring, and wedge represent fundamental mechanical systems.
- The wedge is explicitly labeled as "near Earth's surface," emphasizing gravitational context.
#### Experimental Schematics
- **Diagrams (1â4)**: Focus on linear motion, spring dynamics, and coupled systems.
- **Diagrams (5â9)**: Introduce gravity, inclined planes, and universal gravitation.
- Arrows indicate force directions (e.g., gravity, spring forces, particle interactions).
#### Derived Laws
- **Energy Conservation**: Combines kinetic energy ($ T_\kappa $), spring potential ($ V_k $), gravitational potential ($ V_g $), and universal gravitation ($ V_G $).
- **Newton's Second Law**: Relates acceleration ($ a_\kappa $) to forces from springs, gravity, and universal gravitation.
---
### Key Observations
1. **System Complexity**: Experiments progress from simple systems (single objects) to complex interactions (multiple particles, springs, and gravitational forces).
2. **Mathematical Abstraction**: Equations generalize experimental observations into universal laws (e.g., energy conservation, Newtonian mechanics).
3. **Notation Consistency**: Variables like $ \kappa $ (spatial coordinates) and $ \lambda $ (force types) are rigorously defined.
4. **Experimental Context**: The note on $ \delta_\lambda $ highlights the role of experimental design in isolating specific forces.
---
### Interpretation
This diagram bridges empirical observations and theoretical physics:
- **Physical Systems**: The left section grounds the analysis in tangible objects (sphere, spring, wedge).
- **Experimental Dynamics**: The center section visualizes forces and interactions, serving as a conceptual bridge to abstract laws.
- **General Laws**: The right section formalizes these interactions into equations, demonstrating how specific experiments (e.g., inclined planes, springs) lead to universal principles like energy conservation and Newtonian mechanics.
The inclusion of $ \delta_\lambda $ in Newton's law emphasizes the importance of experimental control in isolating variables. The diagram underscores the iterative process of physics: from observation to hypothesis to mathematical formalism.
</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
## Horizontal Bar Chart: Concept Trial Distribution
### Overview
The image displays a horizontal bar chart comparing the number of trials associated with various concepts. The chart uses color-coded bars with error bars to represent variability. Roman numerals (I-VI) divide the x-axis into six sections, suggesting a categorical or phased structure.
### Components/Axes
- **Y-Axis (Concepts)**: Labels include `F_G`, `G`, `V_G`, `F_g`, `F_k`, `V_g`, `V_k`, `P`, `T`, `a`, `g`, `k`, `v`, `m`.
- **X-Axis (Number of Trials)**: Ranges from 0 to 800, divided into six sections labeled IâVI.
- **Legend**: Located on the right, associating colors with concepts (e.g., dark green for `F_G`, medium green for `G`, etc.).
- **Background**: Light green with darker green vertical lines separating sections IâVI.
### Detailed Analysis
1. **Concept Trial Values** (approximate, with error margins):
- `F_G`: 780 ± 40 (dark green, section VI)
- `G`: 740 ± 30 (medium green, section VI)
- `V_G`: 700 ± 25 (light green, section VI)
- `F_g`: 550 ± 35 (dark green, section V)
- `F_k`: 480 ± 20 (medium green, section IV)
- `V_g`: 450 ± 25 (light green, section IV)
- `V_k`: 420 ± 30 (dark green, section IV)
- `P`: 400 ± 50 (medium green, section IV)
- `T`: 300 ± 15 (light green, section III)
- `a`: 250 ± 10 (dark green, section III)
- `g`: 220 ± 15 (medium green, section II)
- `k`: 150 ± 10 (dark green, section II)
- `v`: 120 ± 5 (medium green, section I)
- `m`: 80 ± 10 (dark green, section I)
2. **Error Bars**:
- Largest variability: `P` (±50).
- Smallest variability: `v` (±5).
3. **Section Distribution**:
- **Section I**: `v` (120), `m` (80).
- **Section II**: `g` (220), `k` (150).
- **Section III**: `T` (300), `a` (250).
- **Section IV**: `F_k` (480), `V_g` (450), `V_k` (420), `P` (400).
- **Section V**: `F_g` (550).
- **Section VI**: `F_G` (780), `G` (740), `V_G` (700).
### Key Observations
- **Trend**: Concepts in later sections (VâVI) have significantly higher trial counts than earlier sections (IâIII).
- **Outliers**: `P` (400 ± 50) has the largest error margin, suggesting high variability.
- **Color Consistency**: All bars match their legend colors (e.g., `F_G` is dark green, `G` is medium green).
### Interpretation
The data suggests a progressive increase in trial counts across sections IâVI, with concepts in later sections (e.g., `F_G`, `G`) dominating in trial volume. The error margins indicate that variability is highest for `P` and lowest for `v`. The Roman numerals likely represent phases or stages, implying a structured progression in trial allocation. The chart may reflect resource allocation, experimental design, or prioritization of concepts over time.
</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/.