In recent years, we have grown accustomed to artificial intelligence being able to predict almost any scenario, from stock market behavior to the structure of proteins. But in science, a good prediction is not the same as understanding. A sophisticated algorithm can identify that an apple is falling and calculate the speed at which it will hit the ground with impressive accuracy, but it cannot explain why it falls or extract from the observation the law of universal gravitation, as Newton supposedly did after the apple landed on his head.
A new Israeli development is now attempting to close precisely this gap, enabling language models to extract the fundamental mathematical equations that drive complex scientific processes. In simple terms, the aim is to get AI to formulate laws of nature for us, just as Newton, Einstein or Galileo did.
In a joint study by researchers from the University of Haifa and the Weizmann Institute of Science, to be presented at the annual conference of the Institute of Electrical and Electronics Engineers (IEEE), the scientists unveiled the MEDA system. Developed under the leadership of Dr. Teddy Lazebnik, with David Krongauz, Arad Zolty and Prof. Eran Segal, the system is based on a multi-agent AI architecture that combines large language models with mathematical search mechanisms.
Instead of operating as a black box that spits out numbers, the system scans scientific papers and literature, translates verbal knowledge into mathematical variables and constraints, and formulates transparent differential equations that can be examined, challenged and tested in laboratory experiments. The study underwent peer review and the system's code was released as open source for use by the scientific community.
In a series of experiments involving 12 dynamic systems of varying complexity, ranging from predator-prey dynamics, population growth and the spread of epidemics to models of neural activity, the spread of fake news and the progression of chronic wounds, the system identified the relevant variables in all of them.
AI applications
(Photo: Getty Images)
An evaluation by a team of five mathematicians with doctoral degrees produced a surprising finding: relying solely on raw data reduced the biological plausibility of the results, with the average score falling from about 0.97 with the full system to about 0.60 in data-driven models. In the case of a chronic wound model, adding data points did produce a perfect fit to the curve, but generated equations that did not reflect the actual medical mechanism.
The Israeli development offers an alternative to the prevailing trend among the world's technology powers. In the United States and Europe, prominent symbolic regression tools such as SINDy focus on processing massive amounts of signals from sensors, but struggle to operate without clear physical constraints and tend to generate equations that lack biological meaning. Meanwhile, major players such as Britain's Google DeepMind have revolutionized structural prediction with AlphaFold, but the output remains locked inside a deep neural network.
In China, too, leading research institutions are focusing on physics-informed neural networks (PINNs), but these use known equations to train networks rather than generating the equations themselves from scientific texts. MEDA's advantage lies in its ability to turn free-form text from medical papers into an organized, explainable mathematical structure — a "white box."
The researchers stress that the system is not intended to replace scientists or provide definitive answers to questions about the universe. Instead, it is designed as a practical research tool that narrows the range of hypotheses and makes it possible to design more targeted and far less expensive laboratory experiments.