#4405 AI "Discovery Machines"

Hybrid AI architecture could turn neuromorphic systems into reliable discovery machines

https://techxplore.com/news/2026-05-hybrid-ai-architecture-neuromorphic-reliable.html

#4405 Summarize https://techxplore.com/news/2026-05-hybrid-ai-architecture-neuromorphic-reliable.html

Summary of the Article:

Researchers at Washington University in St. Louis, led by Professor Shantanu Chakrabartty, have developed a novel hybrid AI architecture designed to build highly reliable "discovery machines." While most common AI models function as inference machines (like ChatGPT) or learning machines, discovery machines are designed to solve exceptionally complex optimization problems with potentially trillions of variables—such as mapping out efficient supply chain routes, managing city traffic grids, or designing optimized chemical compounds.

Key Highlights of the Research:

Core Recipe: Published in Nature Communications, the new approach creates a higher-order neuromorphic Ising machine by combining two essential ingredients:
1. Neuromorphic-Inspired Autoencoding: A technique modeled on human neurobiology that iteratively compresses massive streams of data to make highly accurate pattern predictions.
2. Fowler-Nordheim Annealing: A mechanism derived from quantum mechanics that introduces controlled noise and randomness. This allows the system to efficiently bypass classical computing barriers and "tunnel" directly toward the optimal solution (the "Eureka moment").
Guaranteed Convergence: The primary advantage over standard classical computing approaches and previous higher-order Ising models is its strict mathematical guarantee of convergence. Instead of potentially running indefinitely without a result, the system guarantees that it will successfully find a highly optimized, state-of-the-art solution within a predictable timeframe.

This appears to deal primarily with optimization. Are there AI discovery models applying LLMs or the like to discover new math, physics, chemistry knowledge, and/or invent and improve existing systems?

Yes, absolutely. While specialized optimization hardware (like the neuromorphic Ising machine mentioned in the article) is designed to find the optimal path through a massive set of variables, researchers have increasingly paired Large Language Models (LLMs), reinforcement learning, and advanced generative architectures with verification engines to actively discover new scientific knowledge, formulate physical laws, and invent or self-improve systems.

AI discovery models span several groundbreaking domains:

1. Mathematics and Algorithmic Invention

Using LLMs to generate novel code and math requires safeguarding against "hallucinations." Researchers have solved this by pairing creative language models with strict automated logical evaluators.

DeepMind’s FunSearch: Short for "searching in the function space," this system pairs an LLM (generating computer code) with an automated evaluator that weeds out incorrect logic, feeding the best ideas back into the model in an evolutionary loop. FunSearch made the first verifiable scientific discovery for an open problem in mathematics using an LLM by finding new solutions to the highly complex Cap Set problem. It also invented new, highly efficient heuristic algorithms for real-world logistical problems like bin-packing.
AlphaGeometry: Combines a neural language model (for intuitive idea generation) with a symbolic deduction engine (for logical verification) to discover proofs for complex, Olympiad-level geometry theorems.
AlphaDev and AlphaTensor: DeepMind AI systems that discovered fundamentally new, faster sorting algorithms and matrix multiplication routines—improving foundational code logic that had remained static in computer science for decades.

2. Chemistry and Materials Science

Instead of traditional trial-and-error laboratory synthesis, deep learning architectures are simulating molecular physics to discover entirely new matter and biological functions.

GNoME (Graph Networks for Materials Exploration): DeepMind utilized graph neural networks and active learning loops to discover over 380,000 novel stable inorganic crystals (out of 2.2 million candidates predicted). This effectively expanded humanity’s known stable materials catalog overnight, surfacing new candidates for advanced superconductors, solar cells, and next-generation solid-state battery electrolytes.
AlphaFold 3: Moving beyond its original scope of predicting isolated protein structures, AlphaFold 3 integrates generative diffusion models to simulate complex interactions across virtually all biomolecules—including proteins binding with DNA, RNA, chemical ligands, and ions. It allows chemists to discover "cryptic" binding pockets dynamically, opening up pathways for treating previously undruggable diseases.

3. Physics and System Control

AI models are being deployed to observe complex physical phenomena and reverse-engineer the underlying laws, or invent ways to control highly volatile environments.

Symbolic Regression (e.g., AI Feynman): Deep learning frameworks designed to observe raw data from physical systems and autonomously deduce the fundamental mathematical equations governing them.
Nuclear Fusion Plasma Control: DeepMind collaborated with the Swiss Plasma Center (EPFL) to train deep reinforcement learning models to predict and control the erratic behavior of superheated plasma inside a tokamak fusion reactor. The AI successfully invented new magnetic coil configurations to shape and stabilize the plasma in real-time.

4. End-to-End Automated Scientific Discovery Agents

We are seeing the rise of unified agentic workflows where LLMs act as independent researchers capable of improving systems autonomously.

The AI Scientist (Sakana AI): A fully automated pipeline powered by frontier LLMs paired with agentic tree search. Given a broad domain (like machine learning), the system autonomously reads existing literature, brainstorms novel hypotheses, writes code to execute experiments, analyzes the plotted output data, and formats a complete scientific manuscript ready for peer review.
Self-Improving Coding Agents: Frameworks like SWE-agent and advanced autonomous software engineers utilize LLMs to navigate massive codebases, invent new feature implementations, patch security vulnerabilities, and rewrite underlying system software iteratively.

Summary

The overarching paradigm shift in AI is moving beyond inference (summarizing text or generating an image) toward verifiable discovery. By sandwiching the creative generation capabilities of LLMs between ground-truth physics engines, logic solvers, and code compilers, AI systems are actively serving as automated engines of discovery across the hard sciences.

How can symbolic regression be applied to a bicycle?

Applying symbolic regression to a bicycle is a brilliant way to understand how artificial intelligence can discover the fundamental laws of physics and mechanics from raw observational data.

Unlike standard machine learning models (like neural networks) that act as "black boxes" to predict outcomes, symbolic regression searches for the actual mathematical equations that govern a system.

Here is how you could apply symbolic regression to deconstruct and understand a bicycle across its core mechanical systems:

1. Kinematics and Gearing (Discovering Mechanical Advantage)

Imagine you attach sensors to a bicycle but know absolutely nothing about gears, radii, or physics. You record raw data streams:

Inputs: Pedal cadence ( $\text{RPM}_{\text{pedal}}$ ), front chainring tooth count ( $T_{\text{front}}$ ), rear cassette cog tooth count ( $T_{\text{rear}}$ ), and wheel radius ( $r$ ).
Output: Rear wheel rotational speed ( $\text{RPM}_{\text{wheel}}$ ) and linear speed ( $v$ ).

The AI's Task: You feed this raw data into a symbolic regression engine. The AI combines basic mathematical building blocks ( $+, -, \times, \div$ , variables, and constants) to explain the data.

The Discovery: It would autonomously derive the exact law of mechanical gear ratios:

v = \left( \frac{T_{\text{front}}}{T_{\text{rear}}} \right) \times \text{RPM}_{\text{pedal}} \times 2\pi r

v = \left( \frac{T_{\text{front}}}{T_{\text{rear}}} \right) \times \text{RPM}_{\text{pedal}} \times 2\pi r

2. Dynamics and Power (Deriving the Cycling Power Equation)

If you equip a cyclist with a power meter, a speedometer, an anemometer (for wind speed), and an inclinometer (for hill slope), you can use symbolic regression to discover where the rider's energy is going.

Inputs: Rider power output ( $P$ ), velocity ( $v$ ), wind speed ( $v_{\text{wind}}$ ), total mass ( $m$ ), and road grade ( $\theta$ ).
Output: Acceleration ( $a$ ).

The Discovery: The algorithm would test thousands of mathematical structures against the data. It would eventually isolate the distinct physical forces resisting the rider, outputting an equation revealing:

A term linear to velocity (rolling resistance: $C_{rr} \cdot m \cdot g \cdot v$ )
A term cubic to velocity (aerodynamic drag: $\frac{1}{2} \cdot \rho \cdot C_d A \cdot v_{\text{air}}^3$
A term related to gravity (hill climbing: $m \cdot g \cdot v \cdot \sin(\theta)$ )

3. Stability and Control (The Hardest Problem)

Bicycle stability is notoriously complex; a moving bike stays upright due to a delicate interplay of gyroscopic effects, trail (caster effect), and mass distribution.

By tracking the steering angle, lean angle, velocity, and fork geometry over time, symbolic regression can deduce the differential equations that govern self-stability (how a bike steers itself under falling weight to stay upright), helping engineers optimize frame geometry without endless trial-and-error prototyping.

Explore the Concept: Symbolic Regression Discovery Lab

To see how an AI discovers these formulas real-time, you can use the interactive sandbox below. Adjust the physical parameters of a simulated cyclist, generate "sensor data," and watch how a symbolic regression algorithm builds mathematical trees to uncover the hidden physics of the ride.

How would the cause-effect chain be defined symbolically?

To define a cause-effect chain symbolically, we must move beyond standard algebraic notation. Standard algebra is symmetric: the equation $y = 2x$ logically implies $x = \frac{y}{2}$ .

However, causality is fundamentally asymmetric. If pedaling a bicycle causes the wheels to turn, physically turning the wheels does not cause your legs to pedal.

In mathematics, artificial intelligence, and physics, symbolic cause-effect chains are defined using three primary frameworks: Structural Causal Models, Directed Graphs, and Dynamical Systems.

1. Structural Causal Models (SCMs) & Asymmetric Assignment

Pioneered by computer scientist Judea Pearl, SCMs replace the standard equals sign ( $=$ ) with an asymmetric assignment operator ( $\leftarrow :=$ ).

A causal chain where Variable $X$ causes Variable $Z$ , which in turn causes Variable $Y$ , is defined by a set of structural equations:

X \leftarrow U_X

Z \leftarrow f_Z(X, U_Z)

Y \leftarrow f_Y(Z, U_Y)

Endogenous Variables ( $X, Z, Y$ ): Variables whose values are determined by the model.
Exogenous Noise ( $U_X, U_Z, U_Y$ ): Unobserved external factors or random physical disturbances.
Functions ( $f$ ): The deterministic physical or logical laws mapping the cause to the effect.

2. The Symbolic $do$ -Calculus (Intervention vs. Observation)

To symbolically prove a cause-effect chain, AI must distinguish between passively observing data and actively intervening in the system. Pearl defined the symbolic $do$ -operator to represent this:

Observation ( $P(Y \mid X=x)$ ): "What is the probability of effect $Y$ given that we see $X$ happen?" (Correlation).
Intervention ( $P(Y \mid do(X=x))$ ): "What is the probability of effect $Y$ if we artificially force $X$ to happen?" (Causality).

If $X \rightarrow Y$ is a true causal chain, then $P(Y \mid do(X)) \neq P(Y)$ $P(Y \mid do(X)) \neq P(Y)$ . However, forcing the effect does not change the cause: $P(X \mid do(Y)) = P(X)$ .

3. Graphical Symbolism (Directed Acyclic Graphs - DAGs)

Symbolic algorithms map causal dependencies using nodes and directed arrows. The three fundamental symbolic junctions that govern all cause-effect flow are:

The Chain (Mediation): $X \rightarrow Z \rightarrow Y$
- Example: Rain ( $X$ ) causes a wet road ( $Z$ ), which causes a bicycle to slip ( $Y$ ).
The Fork (Confounding): $X \leftarrow Z \rightarrow Y$
- Example: A steep hill ( $Z$ ) causes both a drop in bicycle speed ( $X$ ) and an increase in rider heart rate ( $Y$ ). $X$ and $Y$ are correlated, but neither causes the other.
The Collider (Selection): $X \rightarrow Z \leftarrow Y$
- Example: High rider fitness ( $X$ ) or a strong tailwind ( $Y$ ) can both independently cause high cycling speeds ( $Z$ ).

4. Dynamical Systems (Time-Evolution Causality)

In physics discovery engines, cause and effect are strictly ordered by time. The state of a system at time $t$ acts as the cause for the state at time $t + dt$ . This is defined symbolically using differential equations:

\frac{d\mathbf{S}}{dt} = \mathcal{F}(\mathbf{S}(t), \mathbf{A}(t))

Where $\mathbf{S}(t)$ is the vector of current states (e.g., position, velocity), $\mathbf{A}(t)$ represents external actions/forces applied, and the derivative $\frac{d\mathbf{S}}{dt}$ is the resulting instantaneous effect.

Explore the Concept: Causal Chain & Intervention Simulator

To understand how symbolic systems model cause-effect chains, explore the interactive simulator below. You can observe how signals naturally flow through a chain ( $X \rightarrow Z \rightarrow Y$ ), and test the mathematical $do$ -operator by intervening on a variable to see how it breaks upstream connections while driving downstream effects.

From <https://gemini.google.com/app> Google Gemini 3.0 (Fast)

Search This Blog

Learning from ChatGPT