Researchers have unveiled Korzhinskii-Net, a new neural-network architecture that fuses classical reactive-transport theory with state-of-the-art coordinate-based deep learning. The system is designed to model the coupled flow, heat, and chemical-reaction processes that govern how hydrothermal fluids redistribute elements through rock — the same processes responsible for the formation of many of the world’s most important ore deposits. The work is led by Boris Kriuk, who serves as research lead on the project.
Where conventional reactive-transport simulators rely on expensive finite-volume or finite-element discretizations and can take hours to days to resolve a single geological scenario, Korzhinskii-Net learns a continuous representation of the underlying fields and evaluates them in milliseconds once trained. Crucially, it does so while respecting the structural constraints that geochemists have known for decades — constraints that purely data-driven surrogates routinely violate.
Encoding a classical theory into a modern network
The architecture takes its name from the mid-twentieth-century theory of infiltration metasomatism, which describes how a moving fluid in chemical disequilibrium with its host rock produces a characteristic sequence of sharply bounded mineral zones. That theory predicts the order in which components are mobilized, the geometry of reaction fronts, and the way zone widths “telescope” with distance from the fluid source. These are not soft empirical regularities; they are consequences of local thermodynamic equilibrium, and any credible simulator must reproduce them.
Korzhinskii-Net builds constraints into the network itself rather than hoping they emerge from data. A coordinate input — a point in two-dimensional space — is first lifted into a high-frequency feature space using random Fourier features, a technique that allows compact networks to represent the sharp gradients found at reaction fronts. A shared multilayer-perceptron trunk then feeds three specialized output heads that emit the temperature, pressure, and concentration fields. A dedicated reaction module composes these fields into a chemical modulator that encodes which reactions are active where, drawing directly on the formal structure of classical metasomatic zoning.
Training is driven by a composite loss that simultaneously penalizes violations of the Darcy flow equation, the heat-transport equation, and the reactive-transport equation, alongside boundary conditions and physically motivated sign constraints that prevent the network from predicting, for example, negative concentrations or thermodynamically forbidden zone sequences.
Why it matters
“The bottleneck in modeling ore-forming systems has never really been a lack of theory — the theory has been in place for sixty years,” Kriuk explained. “The bottleneck is computational. Every time you want to test a new geological scenario, you pay the full cost of a stiff, multi-physics simulation. We wanted an architecture that could amortize that cost while still being recognizably faithful to the underlying chemistry.”
The implications extend well beyond academic geochemistry. Rapid, physically consistent surrogates for reactive transport could accelerate mineral exploration by allowing geologists to screen many more candidate scenarios than is currently feasible; support geothermal energy development, where fluid–rock interaction controls reservoir longevity; and improve risk assessment for carbon sequestration projects, where injected CO₂ reacts with host rock over long timescales. Each of these applications shares the same underlying mathematics, and each is currently rate-limited by simulation cost.
A different philosophy from black-box surrogates
Recent years have seen a proliferation of machine-learning surrogates for subsurface processes, most of them trained on large ensembles of forward simulations. While impressive in their speed, such models tend to treat chemistry as a black box and can produce predictions that are locally plausible but globally inconsistent — a serious problem in a discipline where small thermodynamic violations propagate into large errors in predicted ore-grade distributions.
Korzhinskii-Net takes the opposite approach. Rather than learning the chemistry from scratch, it bakes the known structure of metasomatic systems into the architecture and uses the loss function only to fit the residual degrees of freedom. The result, according to the research team, is a model that requires substantially less training data than purely data-driven alternatives and that fails gracefully — its errors remain within the manifold of physically admissible solutions even when the network is queried far from its training distribution.
Looking ahead
The team plans to extend Korzhinskii-Net to three-dimensional domains, to richer multi-component chemical systems, and eventually to inverse problems in which observed mineral assemblages are used to reconstruct the fluid history that produced them. Kriuk emphasized that the broader goal is methodological: “We see this as a template. Many natural sciences have well-developed classical theories that modern deep learning has so far ignored. The interesting question is how to put them back in.”
The Korzhinskii-Net architecture and accompanying benchmarks will be described in a forthcoming technical manuscript.
