Reactive Transport Modeling with Physics-Informed Machine Learning for Critical Minerals Applications

This study introduces a physics-informed neural network (PINN) framework designed to accurately simulate fast bimolecular reactions in porous media, thereby enhancing the characterization of chemical interactions essential for critical mineral extraction and geoscience applications.

The Gist

Imagine you are trying to predict how a drop of dye spreads through a sponge that is soaked with water, but with a twist: as the dye moves, it instantly reacts with a second liquid to create a brand new color. Now, imagine that sponge isn't uniform; some parts are soft and spongy, while others are hard and rocky, causing the water to swirl, speed up, or get stuck in unexpected places.

This is the real-world problem scientists face when trying to extract critical minerals (like lithium or rare earth elements used in our phones and electric cars) from deep underground. They need to pump chemicals into the ground to dissolve the minerals, but they need to know exactly where the chemicals will go and how they will mix to avoid wasting money or creating pollution.

Traditionally, solving this puzzle requires massive, complex computer simulations that are slow, expensive, and sometimes make mistakes (like predicting negative amounts of a chemical, which is physically impossible).

This paper introduces a new, smarter way to solve this puzzle using "Physics-Informed Machine Learning" (PINNs).

Here is the breakdown of their approach using simple analogies:

1. The Old Way vs. The New Way

• The Old Way (Traditional Simulations): Imagine trying to map a city by drawing a grid of tiny squares on a piece of paper and calculating the traffic flow for every single square. If the city has weird shapes or sudden changes in road quality, you have to redraw the grid thousands of times. It's tedious, requires a lot of paper (computing power), and if you miss a square, your map is wrong.
• The New Way (PINNs): Imagine you have a super-smart GPS that doesn't need a grid. Instead, it "knows" the laws of physics (like "cars can't drive through buildings" or "traffic flows downhill"). You just tell it the starting point and the destination, and it learns the path by constantly checking its own predictions against the laws of physics. It doesn't need a grid; it just needs to know the rules.

2. The Three-Step Recipe

The authors built a system that solves the problem in three stages, like a cooking recipe:

• Step 1: The Flow (The Water Current)
  First, they teach the AI how water moves through the rocky underground. They use a "mixed" approach, which is like making sure the AI understands both the pressure pushing the water and the speed of the water itself. They tested this on "patch tests" (simple scenarios with different rock types) and proved the AI could predict the water flow just as accurately as the old, heavy-duty methods.

• Step 2: The Spread (The Diffusion)
  Next, they teach the AI how chemicals spread out. A major problem with old computer models is that they sometimes predict "negative" amounts of chemicals (like saying you have -5 apples), which makes no sense.

  • The Magic Trick: The PINN method naturally respects the "Maximum Principle." Think of it like a strict teacher who ensures no student ever gets a negative score. The AI learned to keep all chemical concentrations positive without needing extra rules, something the old methods struggled to do.

• Step 3: The Reaction (The Magic Mix)
  Finally, they simulated the fast reaction. Imagine two streams of water meeting: one with Acid (A) and one with a Complexing Agent (B). When they touch, they instantly turn into a Metal Complex (C).

  • The Challenge: In the real world, if the water is moving fast and the rocks are uneven, the mixing happens in thin, sharp lines. Old computers often blur these lines, making the reaction look fuzzy.
  • The Result: The PINN method kept the lines sharp and clear. It showed exactly where the "plume" (the cloud of the new chemical) would form, even when the water was swirling randomly.

3. Why This Matters for Critical Minerals

Extracting minerals from the ground is like trying to find a needle in a haystack while blindfolded. You pump chemicals in, hoping they hit the minerals.

• Data Scarcity: We often don't have enough data from underground to train normal AI.
• The PINN Advantage: Because this AI "knows" the laws of physics, it doesn't need a mountain of historical data. It can learn from the physics itself. This means engineers can simulate scenarios, optimize how much chemical to inject, and predict where the minerals will be recovered, all without needing expensive, slow supercomputers.

The Bottom Line

The authors have created a digital twin for underground chemical reactions. It's a tool that is:

1. Mesh-free: No need to draw complex grids.
2. Physically Honest: It never predicts impossible results (like negative chemicals).
3. Fast and Flexible: It can handle messy, real-world underground conditions where water flows in unpredictable ways.

This technology could be a game-changer for the green energy transition, helping us mine the critical minerals we need for batteries and electronics more efficiently and with less environmental impact.

Technical Summary

Here is a detailed technical summary of the paper "Reactive Transport Modeling with Physics-Informed Machine Learning for Critical Minerals Applications."

1. Problem Statement

The paper addresses the challenges of modeling reactive transport in heterogeneous porous media, specifically focusing on fast bimolecular reactions relevant to critical mineral extraction (e.g., in-situ leaching).

• The Challenge: In subsurface environments, chemical reactions often occur on timescales much faster than transport (advection/diffusion). This leads to "mixing-limited" reactions where the reaction rate is controlled by the incomplete mixing of reactants rather than intrinsic kinetics.
• Numerical Difficulties: Traditional numerical methods (Finite Element Method - FEM, Finite Volume Method - FVM) struggle with these problems due to:
  • Numerical Diffusion: Smearing out sharp reaction fronts.
  • Non-negativity Violations: Producing unphysical negative concentrations, violating the maximum principle.
  • Data Scarcity: Subsurface applications often lack the large datasets required for purely data-driven machine learning models.
• Goal: To develop a robust, mesh-free framework capable of accurately predicting sharp reaction fronts and ensuring physical constraints (like non-negativity) without relying on massive training datasets.

2. Methodology

The authors propose a sequential Physics-Informed Neural Network (PINN) framework that decouples the problem into three stages: Flow, Transport, and Reaction.

A. PINN Framework

• Architecture: A Deep Neural Network (DNN) using the tanh activation function.
• Training Strategy: The network is trained by minimizing a composite loss function ($L$) consisting of:
  • PDE Residual Loss ($L_{PDE}$): Enforcing governing equations at collocation points within the domain.
  • Boundary Condition Loss ($L_{BC}$): Enforcing Dirichlet/Neumann conditions.
  • Weighting: Adaptive weighting ($\lambda_1, \lambda_2$) is used to balance the contribution of different loss terms.
• Solver: The framework utilizes the JAX library for automatic differentiation and the Adam optimizer.

B. Sequential Modeling Stages

1. Flow Subproblem (Darcy Flow):

   • Solved using a stabilized mixed formulation (simultaneously solving for pressure $p$ and velocity $v$).
   • Why Mixed? Single-field formulations often yield poor velocity accuracy. The mixed formulation ensures flux continuity, which is critical for heterogeneous media.
   • Output: A velocity field $v(x)$ used to compute the dispersion tensor.

2. Transport Subproblem (Anisotropic Diffusion):

   • Modeled using an anisotropic diffusion equation where advection is implicitly handled via a velocity-dependent mechanical dispersion tensor $D(x)$.
   • Key Feature: The study explicitly tests the Maximum Principle (ensuring concentrations remain non-negative).

3. Reaction Subproblem (Fast Bimolecular Reaction):

   • Reaction Type: $n_A A + n_B B \rightarrow n_C C$.
   • Assumption: The reaction is instantaneous (fast limit), meaning reactants $A$ and $B$ cannot coexist at the same point ($c_A \cdot c_B = 0$).
   • Innovation (Invariant Reformulation): Instead of solving stiff coupled reaction-diffusion equations, the authors reformulate the system using chemical invariants ($\Psi_A, \Psi_B$):
     • $\Psi_A = c_A + \frac{n_A}{n_C}c_C$
     • $\Psi_B = c_B + \frac{n_B}{n_C}c_C$
   • These invariants satisfy uncoupled diffusion equations ($-\text{div}[D \nabla \Psi] = 0$).
   • Reconstruction: After training the PINNs to solve for $\Psi_A$ and $\Psi_B$, the actual species concentrations ($c_A, c_B, c_C$) are recovered algebraically using stoichiometry and non-negativity constraints (e.g., $c_A = \max(\dots, 0)$).

3. Key Contributions

1. Invariant-Based PINN Formulation: The paper introduces a novel approach to solving fast bimolecular reactions by decoupling the reaction source terms via chemical invariants. This avoids the stiffness associated with solving coupled reaction-diffusion equations directly.
2. Validation of Non-Negativity: The study demonstrates that PINNs inherently satisfy the maximum principle (non-negative concentrations) in diffusion problems, whereas traditional FEM often produces spurious negative values without complex stabilization techniques.
3. Mixed-Formulation Flow Coupling: The successful integration of a mixed-formulation PINN for flow ensures accurate velocity fields in heterogeneous media, which is a prerequisite for accurate reactive transport modeling.
4. Mesh-Free Flexibility: The framework operates without a physical mesh, utilizing collocation points, which simplifies handling complex geometries and heterogeneous properties.

4. Results

The framework was validated through three progressively complex benchmarks:

• Patch Tests (Flow Verification):

  • Three tests (Vertical, Horizontal, Inclined) with heterogeneous permeability ($k_1=1, k_2=10$) were conducted.
  • Result: PINN predictions for pressure and velocity matched FEM solutions (COMSOL) with high accuracy, validating the mixed formulation's ability to handle permeability discontinuities.

• Transport Problem (Maximum Principle):

  • An anisotropic diffusion problem with a central hole was solved.
  • Result: PINNs produced strictly non-negative concentrations. In contrast, the FEM solution exhibited negative concentrations (violating the maximum principle) in regions of high anisotropy, requiring additional constraints to fix.

• Fast Bimolecular Reactions:

  • Uniform Flow: The model successfully captured the sharp mixing front between reactants A and B and the formation of product C.
  • Non-Uniform (Random) Flow: The model handled a complex, explicit velocity field with random variations.
  • Result: In both cases, PINNs accurately predicted the "plume" of product C, capturing the sharp, mixing-limited reaction fronts. The results closely matched reference FEM solutions from previous literature (Mudunuru and Nakshatrala, 2016).

5. Significance and Applications

• Critical Minerals: The methodology is directly applicable to in-situ leaching and critical mineral recovery (e.g., Rare Earth Elements), where optimizing reagent injection and minimizing by-products depend on accurately predicting mixing-limited reaction fronts.
• Data Efficiency: Unlike data-driven ML, this approach requires no experimental training data, making it ideal for subsurface applications where data is scarce or expensive to obtain.
• Physical Consistency: By embedding physics laws directly into the loss function, the model guarantees adherence to conservation laws and physical constraints (like non-negativity) that are often violated by standard numerical or data-driven methods.
• Future Potential: The framework offers a scalable path for 3D simulations, inverse modeling, and real-time monitoring of subsurface chemical processes, potentially replacing computationally expensive mesh-based solvers in specific scenarios.

In conclusion, the paper establishes that PINNs, when combined with chemical invariants and mixed-flow formulations, provide a robust, mesh-free, and physically consistent tool for simulating complex reactive transport phenomena in critical mineral applications.