Meet Praadhyumn Indaana, the 17-year-old Montvale, New Jersey student whose AI model cut nuclear reactor heat-limit prediction error from about 63% to 5.5%; he won $50,000
A better understanding of how much heat a nuclear reactor can safely handle could have implications far beyond a single engineering calculation. For 17-year-old Praadhyumn Indaana of Montvale, New Jersey, that challenge became the focus of an eight-month research project combining physics and artificial intelligence. According to the Davidson Institute, which named Indaana a 2026 Davidson Fellow, his physics-guided neural network reduced the average prediction error for critical heat flux from about 63% with conventional formulas to 5.5%. The work earned him a $50,000 Davidson Fellows Scholarship and recognition for a project aimed at improving predictions of an important nuclear-reactor operating limit.
Why heat matters inside a nuclear reactor
In a nuclear reactor, water plays a critical role in removing heat from fuel rods. As the water boils against the fuel rods, it carries heat away. But there is a point at which the boiling process can no longer remove heat effectively. This threshold is known as critical heat flux, or CHF. Going beyond this limit can cause the cooling process to deteriorate rapidly. Steam can displace liquid water from the surface of the fuel rod, reducing cooling and causing the rod temperature to rise sharply. Because of the potential safety consequences, reactor operators and engineers need reliable ways to predict where this limit occurs.The difficulty, Indaana explains in his Davidson Institute project, is that conventional equations have generally been developed and tuned around particular experimental conditions. Their accuracy can therefore decline when they are applied outside those conditions.
Building an AI model with physics built in
Indaana approached the problem using a physics-regularized neural network. Instead of allowing an artificial-intelligence model to learn patterns from experimental data alone, he incorporated constraints based on physical theory. The model was trained using data from more than 10,000 real experiments. Indaana also incorporated bounds derived from hydrodynamic instability theory, helping prevent the model from producing predictions that violate known physical limits.This distinction was important because a conventional machine-learning model can struggle when it encounters situations for which there is little training data. According to Indaana, early versions of his models performed well in regions with large amounts of data but could produce physically unreasonable predictions when extrapolating into less represented high-quality-flow regions. He addressed the problem by adjusting the physics-based regularisation according to how closely different empirical correlations agreed for individual samples. He also added a penalty when predictions exceeded a known physical upper bound.
Error falls from about 63% to 5.5%
The resulting model produced a major improvement in prediction accuracy. According to the Davidson Institute, the conventional formulas showed average errors of about 63%, while Indaana’s model reduced that figure to 5.5%. On a fixed testing dataset, the model achieved an R² value of 0.986, a statistical measure indicating how closely the model’s predictions matched the observed data in that evaluation. Indaana compared its performance with widely used correlations, which he says produced errors an order of magnitude larger.The achievement does not mean that nuclear reactors can simply be operated closer to the heat limit. Reactor safety requirements remain stringent. Instead, more accurate prediction could help engineers understand the boundary more precisely and potentially reduce some of the conservatism built into reactor design and operation.
An idea that began with a question about fusion
The project originated during a class on fusion energy that Indaana took through the Columbia University Science Honors Program. He asked why reactors that are already well characterised still need to operate conservatively. The answer led him to critical heat flux and the uncertainty surrounding its prediction across different operating regimes. He subsequently decided to investigate whether machine learning could improve the situation while remaining constrained by established physical knowledge.His interests extend beyond conventional nuclear reactors. Indaana told the Davidson Institute that he is interested in the intersection of materials science, quantum physics, computing and nuclear fusion, and hopes to study physics in college. His project also explores how physics-informed machine learning can be applied to engineering problems where knowledge is often expressed through empirical correlations and physical bounds rather than complete governing equations.
An independent research effort
Indaana completed the project over roughly eight months, working through literature review, data collection, feature engineering, model development, statistical evaluation and paper writing. He taught himself much of the thermohydraulics, physics-informed machine learning and statistics required for the work.He also used statistical methods including bootstrap resampling and paired nonparametric tests rather than relying solely on a reduction in prediction error. The model development and experiments were carried out on his own computer, with cloud resources such as Kaggle used for training. He received guidance from a mentor on the broader research process.
Recognition with a $50,000 scholarship
Indaana’s work earned him a place among the 2026 Davidson Fellows, a programme recognising students aged 18 and younger for significant achievements in science, technology, engineering, mathematics, literature and music. The Davidson Institute awarded him a $50,000 scholarship for his project.His research illustrates how artificial intelligence can be used not simply to process large amounts of data, but to combine data-driven learning with established scientific principles. In Indaana’s case, that approach produced a substantially more accurate way of predicting a difficult engineering boundary, potentially offering a new way to approach problems where safety depends on understanding the limits of physical systems.
