The alternative is known as inline, or tight, coupling. Instead of fully converging the neutronics solution before passing information across, the solver pauses partway through, hands a partially converged power distribution to the thermal-hydraulics code, receives updated temperature and density fields, and resumes where it left off. Theoretical work and practical experience have shown that this approach can make the coupled calculation unconditionally stable while dramatically reducing runtime, because only a small fraction of a full neutronics solve is needed between physics updates. What has remained murky is how the statistical noise inherent to Monte Carlo neutronics interacts with this frequent exchange of data, and whether the choices an analyst makes, such as the number of neutrons simulated per cycle or how often the physics fields are refreshed, might undermine the whole enterprise.

That question has now been addressed by P. Cosgrove of the University of Cambridge, using newly developed multi-physics capabilities in the open Monte Carlo code SCONE. The work, published on 17 September 2026, reports a series of numerical experiments on a three-dimensional pressurised water reactor fuel pin, a benchmark comparable to those used in previous studies of coupled Monte Carlo and thermal-hydraulics. The pin, with 2.2 percent enriched fuel, a height of 365.76 centimetres, and water coolant at 2215 psi, was discretised into 36 axial zones. Thermal-hydraulics was handled by a simple enthalpy-balance script that updated coolant temperature and density along those zones and computed fuel surface and centreline temperatures, while SCONE tracked neutrons through an imposed, piecewise-constant temperature and density field using ENDF/B-VIII.0 nuclear data.

The implementation itself is a notable piece of engineering. SCONE now accepts temperature and density fields superimposed on the geometry, and rather than relying on rejection sampling to track particles through changing fields, it explicitly computes the distance at which the field value changes and moves the neutron there if that distance is shorter than the distance to collision. A new ‘coupling admin’ object manages all communication between the Monte Carlo transport, the tallying, and external coupling scripts, controlling how often physics updates occur, which tallies are exchanged, and whether coupling continues during the tallying phase of the calculation. Doppler broadening in the fuel is handled on the fly by target motion sampling, and moderator temperatures by stochastic interpolation, so the code can respond to feedback without reloading nuclear data.

The experiments compared inline coupling against traditional under-relaxed Picard iteration, which served both as the source of a high-fidelity reference solution and as a baseline for common practice. The reference calculation used two million neutrons per cycle, 200 inactive cycles, and 200 active cycles, repeated ten times independently. The inline calculations then varied three parameters: the frequency of physics updates measured in Monte Carlo cycles, the fraction by which the neutron population was scaled relative to the reference, and whether coupling was continued into the active cycles. The total number of active neutrons was held constant across all runs, and each configuration was repeated five times with different random number seeds.

The Picard results first illustrate why population size matters so much. With the same total number of active neutrons, reducing the population per cycle from one million to twenty thousand increased the maximum power error from just over one percent to about four percent, and the maximum fuel temperature error from roughly 0.4 percent to about one percent. The reason is inter-generational correlation: in a Monte Carlo criticality calculation, successive generations of neutrons are not statistically independent, because each generation’s fission source is sampled from the previous one. With fewer neutrons per generation, this correlation produces a higher true variance than nominal error estimates suggest, and that hidden variance becomes a real bias once the noisy power distribution is propagated through the coupled physics.

The inline results delivered the study’s headline surprises. First, continuing coupling into the active cycles is essential. With active coupling, power errors of order one percent and temperature errors of order one tenth of a percent were achieved across all population sizes and update frequencies, and the method proved robust to those choices. Without it, errors could deteriorate by an order of magnitude, because fixing the thermal-hydraulic fields based on noisy early tallies locks in a bias. Second, and unexpectedly, inline coupling with active updates was markedly more accurate than Picard iteration at the same neutron population per generation, and roughly ten times faster, since only one transport solve was needed instead of ten. The advantage faded at large populations, which pointed the investigators toward a statistical explanation.

That explanation is the most intriguing finding of the work. In a Picard iteration, the fission source evolves over many cycles against a fixed thermal-hydraulic field, allowing inter-generational correlation to build up largely unimpeded. Under inline coupling, the temperature and density fields are themselves repeatedly modified by the noisy power tallies, and Cosgrove hypothesises that this continual adjustment disrupts the accumulation of correlation between neutron generations, perhaps because the physics feedback keeps nudging the neutron distribution toward the coupled fundamental mode. To test the idea, autocorrelation coefficients of the spatially averaged power production were computed over ten independent runs, both with a fixed field and with inline coupling updating every cycle at twenty thousand neutrons per cycle. The coupled case showed clearly lower correlation across all lags, and the measured root-mean-square error agreed well with theoretical predictions derived from the autocorrelation analysis.

The practical consequence is striking: even when the correct temperature and density fields are already known, running the coupled inline calculation yields more accurate power tallies than an uncoupled one for the same number of active particles. For the fuel pin problem, the analysis implies that about 2.89 times fewer active particles need to be simulated in the coupled calculation to achieve the same power errors, with the reduction potentially greater in practice. Whether the trade is worthwhile will depend on the cost of the thermal-hydraulics solver and the data exchange; inline coupling only pays off when neutronics dominates the computational budget, which would not be the case with, say, a full computational fluid dynamics solver on the other side of the interface.

The study closes with caveats and a research agenda. The extent of correlation reduction will likely depend on the strength and sign of the reactivity feedback, the size of the problem, and the geometry: thermal-hydraulic coupling strongly shapes the axial power profile, but realistic full-core problems are much larger radially, and it is unclear whether the same benefit carries over. The findings do, however, lend new support to existing inline practices, such as the xenon equilibrium iterations long used in codes like Serpent and MC21, which this work effectively justifies continuing through active cycles. Future work will apply the new SCONE coupling machinery to larger systems with higher-fidelity solvers, and develop a stochastic theory of how multi-physics feedback accelerates statistical convergence. For now, the message to reactor analysts is clear and a little counterintuitive: letting the physics talk more often does not just stabilise the calculation, it quietly cleans up the noise.

Subject of Research: Inline multi-physics coupling of Monte Carlo neutronics and thermal-hydraulics in reactor simulations

Article Title: Better physics, less noise: inline multi-physics reduces correlation in Monte Carlo reactor simulations

Article References: Cosgrove, P. (2026). Better physics, less noise: inline multi-physics reduces correlation in Monte Carlo reactor simulations. Frontiers in Nuclear Engineering, 5, Article 1921592. https://doi.org/10.3389/fnuen.2026.1921592

Image Credits: AI Generated

DOI: 10.3389/fnuen.2026.1921592

Keywords: Monte Carlo, neutronics, thermal-hydraulics, multi-physics coupling, SCONE, inter-generational correlation, Picard iteration, reactor simulation, fuel pin, variance reduction, nuclear engineering, PWR