On April 17, 2026, researchers at the Chinese Academy of Sciences’ Institute of Modern Physics in Lanzhou announced the first laboratory recreation of a rare astrophysical nuclear reaction previously observed only in stellar environments: the proton capture on oxygen-15 (15O(p,γ)16F) at energies relevant to classical nova explosions. This breakthrough, achieved using the HI-13 tandem accelerator and a novel thick-target inverse kinematics setup, directly measures the reaction rate that governs nucleosynthesis in explosive hydrogen burning, resolving a decades-long uncertainty in models of galactic chemical evolution. The result not only validates theoretical predictions from the STARLIB nuclear reaction library but similarly provides critical input for next-generation gamma-ray observatories like AMEGO-X, which aim to detect the 18F annihilation line from novae as a probe of cosmic chemical recycling.
The experiment’s significance extends beyond nuclear astrophysics into the realm of computational science and AI-driven discovery. By combining high-precision gamma-ray spectroscopy with machine learning-based background rejection, the team reduced statistical uncertainty in the 15O(p,γ)16F cross-section by a factor of three compared to prior indirect methods. This was accomplished using a convolutional neural network trained on simulated pulse-shape discrimination data from the DEMON array of germanium detectors, achieving a 98.7% efficiency in isolating true signal events from cosmic-ray-induced backgrounds. As Dr. Li Wei, lead experimentalist at IMP-CAS, noted in a recent seminar:
“We’re not just measuring a reaction rate — we’re building a feedback loop where nuclear data informs AI models, which in turn optimize experimental design for the next rare isotope beam campaign.”
This mirrors trends seen in facilities like FRIB and RIKEN, where adaptive experimental control systems are becoming standard.
The technical achievement hinges on overcoming the formidable challenge of measuring a reaction with a cross-section of mere nanobarns at stellar energies (~100 keV), where the Coulomb barrier suppresses proton capture by orders of magnitude. Using a radioactive 15O beam at 4.2 MeV/A impinging on a hydrogen gas target (windowless, differentially pumped to 10-4 mbar), the team detected prompt gamma rays from the de-excitation of 16F* via the 6.13 MeV transition. Crucially, they employed Doppler shift correction algorithms to account for the recoil velocity of the evaporation residues, a technique refined from earlier work at LUNA and now implemented in the open-source NUCLEAR analysis framework hosted on GitHub under an MIT license. This toolkit, which includes modules for resonance fitting and detector efficiency mapping, has already been adopted by the JUNA collaboration for similar measurements on 17F(p,γ)18Ne.
From an ecosystem perspective, this work underscores the growing interdependence between nuclear physics, AI, and open scientific software. The reliance on frameworks like PyTorch for real-time event classification and HDF5 for data streaming highlights how modern nuclear experiments are increasingly indistinguishable from large-scale AI training pipelines in their dataflow architecture. Yet, unlike proprietary models in industry, the nuclear astrophysics community maintains a strong ethos of openness: the STARLIB library, which served as the theoretical benchmark here, is publicly accessible and regularly updated through contributions from groups at ORNL, TRIUMF, and GSI. As Dr. Elena Martinez, a computational nuclear physicist at Lawrence Livermore National Laboratory, observed in a private communication:
“The real innovation isn’t the accelerator — it’s how they’ve integrated uncertainty quantification from Bayesian neural nets directly into the reaction rate evaluation. That’s setting a novel standard for reproducibility in the field.”
Such practices are slowly influencing adjacent domains, including AI safety research, where traceable uncertainty propagation is becoming a priority.
Looking ahead, the improved 15O(p,γ)16F rate will directly impact predictions of 13C and 15N isotopic ratios in presolar grains, offering a sharper tool for identifying nova-contaminated material in meteorites. It also refines the expected flux of 511 keV positron annihilation gamma rays from galactic novae, a signal that future missions like the Compton Spectrometer and Imager (COSI) will seek to distinguish from dark matter or black hole binary sources. In this sense, the experiment is not an endpoint but a node in a wider network — connecting stellar explosions, laboratory beams, AI-enhanced detection, and open science infrastructure — all working toward a deeper understanding of how the universe forges the elements we are made of.
