The Collision Operator as a Small Circuit: A Conserving Replacement for the Constant Relaxation Time

To predict how fast electrons move through a new semiconductor, physicists solve the Boltzmann transport equation. This equation tracks how many carriers sit in each momentum state and how scattering shuffles them between states. The usual shortcut, the constant relaxation time approximation (CRTA), replaces all of that shuffling with one number $\tau$: every state leaks back to equilibrium at the same rate. The shortcut is cheap, but it is known to fail. It fixes the Hall factor at 1, it misses inelastic phonon scattering, and it breaks particle-number conservation.

This note starts from a fact that is exact but rarely exploited. Once the Boltzmann equation is linearized, it is literally Kirchhoff's current law on a resistor network whose nodes are electron states. In that picture CRTA is the cartoon in which every node is wired to ground through an identical resistor, and the true collision operator is an ungrounded network with resistors between states. We propose to treat it as a circuit: coarsen the network with the structure-preserving reduction methods electronic design automation already uses for interconnect, and keep a 5–20 node circuit that conserves particles, carries magnetic-field physics as non-reciprocal edges, and extrapolates in temperature through explicit phonon-occupation factors. The result is a drop-in replacement for $\tau$ that is about as cheap as CRTA, close to as accurate as a full iterative solve, and natively a netlist.

Core Claim

Write the deviation of the carrier distribution from equilibrium as $-\varphi_k\,\partial f_0/\partial E_k$, where $f_0$ is the Fermi-Dirac distribution and $\varphi_k$ is a "potential" on state $k$. The linearized steady-state Boltzmann equation in an electric field $\mathbf{E}$ then reads

$$\sum_{k'} G_{kk'}\,(\varphi_k-\varphi_{k'}) \;=\; e\,\mathbf{E}\cdot\mathbf{v}_k\left(-\frac{\partial f_0}{\partial E_k}\right) \equiv s_k .$$

Each symbol has a direct circuit reading. $G_{kk'} = f_0(E_k)\,[1-f_0(E_{k'})]\,P_{kk'}/k_BT$ is a conductance between states $k$ and $k'$, built from the transition rate $P_{kk'}$, and it is symmetric by detailed balance. $\mathbf{v}_k$ is the group velocity. $s_k$ is a current injected at node $k$. Because the network has no ground, total current is conserved, which is particle-number conservation.

CRTA replaces the left side by $\varphi_k(-\partial f_0/\partial E_k)/\tau$: no edges, every node grounded. An energy-dependent relaxation time keeps the grounding but lets the resistor values vary. Rode's iterative method, used by Chakrabarty et al. on AlGaAs₂ and by Mandia et al. on Ti₂CO₂, effectively solves the full network. The edges it keeps are the "in-scattering" terms through which polar optical phonons move carriers between energies separated by $\hbar\omega$.

Three more ingredients complete the circuit. A magnetic field adds an antisymmetric term $\mathbf{v}\times\mathbf{B}\cdot\nabla_k$, which acts as non-reciprocal edges along constant-energy orbits; this is what makes the Hall factor differ from 1. Adding node capacitances $C_k=-\partial f_0/\partial E_k$ to ground gives the frequency-dependent conductivity $\sigma(\omega)$. CRTA's single Drude pole is then the one-node limit.

The claim is that a reduced network $G_{\text{red}} = P^\top G P$ is enough. Here $P$ aggregates states into $K$ groups by valley, energy shell and, under a field, orbit sector. This projection keeps symmetry, passivity and the conservation null vector. Its edges are tagged by scattering mechanism (acoustic, piezoelectric, polar optical, impurity, remote substrate phonon), and each tag carries its analytic Bose-occupation temperature dependence. The prediction is that such a network reproduces full-solve mobility, Hall factor and $\sigma(\omega)$ to within a few percent with $K \lesssim 20$, which neither CRTA nor energy-dependent relaxation times can do in inelastic regimes. It should also extrapolate across temperature from a single reduction, because temperature enters through physical occupation factors rather than fitted coefficients.

Relation to Prior Work

The case against CRTA is already established, and these papers motivate the work rather than supply the method. Mandia et al. solve coupled electric and magnetic perturbations beyond relaxation-time models for Ti₂CO₂. They find Hall factors from 0.2 to 1.3, mobility exponents $\gamma$ (in $\mu \propto T^{-\gamma}$) falling from 1.4 to 0.7 with carrier density, and a piezoelectric channel that CRTA hides. The MoS₂ magneto-transport study reports a room-temperature Hall factor near 1.43 for suspended monolayers and shows that substrate remote phonons dominate supported devices. Chakrabarty et al. show that relaxation-time treatments underestimate mobility relative to Rode's method because they mishandle polar optical phonons. Atwal and Ashcroft show that ad-hoc relaxation times violate conservation and corrupt collective response, which is why conservation is a hard design constraint for us.

All of these papers either accept CRTA or pay for a full solve. None produces a compact, conserving intermediate model.

Backman, Lee and Luisier unify linearized Boltzmann and quantum (non-equilibrium Green's function) device transport to within 15%. Their cost was severe enough that they had to restrict phonons to the zone center. Their mobilities (MoS₂ 221 cm²/Vs, silicon 1452 cm²/Vs) serve as our external reference points.

The circuit side supplies the lens and one key piece of evidence. Hakhamaneshi et al. represent circuits with explicit terminal nodes. A graph network trained on Kirchhoff-law DC voltages generalizes zero-shot from 10-branch to 128-branch resistor ladders, which suggests network-structured operators can transfer across sizes. Circuit graph models such as CktGNN and FusionCell parameterize transistors only by geometry and compact-model cards. FusionCell names multi-corner temperature and voltage prediction as unsolved.

We do not propose putting $\tau(E,\mathbf{k})$ into circuit graph networks as node features, which is the obvious move. The reason is that a compact model fitted to Boltzmann mobilities would likely do as well at the circuit level. We instead make the collision operator itself the circuit.

Research Program

We will build a Boltzmann solver that assembles $G$, the magnetic non-reciprocal edges and the node capacitances on a k-space grid. Scattering rates come from the DFT-derived parameters the source papers report (deformation potentials, piezoelectric and polar constants, dielectric functions, phonon energies). The test set is Ti₂CO₂, monolayer MoS₂ suspended and on SiO₂, Al₂O₃ and HfO₂, AlGaAs₂, and bulk silicon. We will validate the solver by reproducing the published mobilities and Hall factors before any reduction.

We then study three ways to choose the aggregation $P$:

  1. fixed physical bins in valley and energy shell, spaced by the optical phonon energy;
  2. spectral clustering on $G$;
  3. Krylov moment matching around $\omega=0$, in the spirit of passivity-preserving interconnect reduction, which matches low-order moments of $\sigma(\omega)$.

Each reduced edge carries its mechanism tag. Temperature transfer applies the analytic occupation and Fermi-factor scaling per tag without re-solving the full network.

Finally, we export each reduced network as a SPICE subcircuit. One port is driven by $\mathbf{E}$ and another reads out current. This shows that the transport model can sit inside an ordinary circuit graph as a terminal-explicit device subgraph.

Baselines

The naive reference is CRTA, with $\tau$ fitted to the full-solve mobility at 300 K for each material and carrier density. This is how BoltzTraP-style workflows are used in practice. It forces the Hall factor to 1 and gives a single-pole $\sigma(\omega)$. Beating it on mobility at the fit point is trivial, so the comparison counts only off the fit point and on Hall factor and $\sigma(\omega)$.

The strongest cheap competitor is an energy-dependent momentum relaxation time combined across mechanisms with Matthiessen's rule. It costs about the same as our reduced network and is the standard improvement over CRTA. It is the main threat wherever scattering is nearly elastic.

The strongest accurate competitor is the full iterative solve, which also serves as ground truth. Its own published outputs anchor the comparison: Hall factor 0.2–1.3 and $\gamma$ from 1.4 to 0.7 for Ti₂CO₂ (Mandia et al.), a Hall factor near 1.43 for suspended MoS₂ (MoS₂ study), and the phonon-limited MoS₂ mobility of 221 cm²/Vs (Backman et al.). For the relaxation-time-versus-Rode gap in AlGaAs₂, no reported numerical result was found on Lacuna; we will measure it ourselves.

The empirical baseline is the compact-model style power law $\mu = AT^{-\gamma}$, with a matching Hall-factor fit. It is fitted on 250–350 K, the same temperature window our reduction sees.

The learned baseline is a small neural network surrogate that maps temperature, density and frequency to mobility, Hall factor and $\sigma(\omega)$. It is trained on as many full-solve calls as our reduction consumes.

The key ablation keeps our aggregation $P$ but deletes all inter-node edges and grounds each node with its reduced diagonal. This keeps the same $K$ and parameter count and removes only in-scattering and conservation.

All methods get the same scattering inputs, the same grid and the same budget of full solves. We count a win only if, at equal $K$, the network is within 5% of the full solve on mobility and Hall factor across 77–500 K and $10^{11}$–$10^{14}$ cm⁻² where the baselines are not.

Why not just…?

Why not just use an energy-dependent relaxation time? It is cheaper to explain and is often adequate. It does, however, lack the inter-energy edges that carry polar optical in-scattering, which is exactly the effect Rode's method corrects in Chakrabarty et al.. It also violates conservation, per Atwal and Ashcroft. We expect it to tie our network in acoustic-dominated, low-temperature regimes, and we will report that honestly. The diagonal ablation identifies where the edges matter: above roughly the temperature where polar optical modes become occupied, and in Hall factor generally.

Why not just run the full iterative solve every time? For a single material at one condition, the solve is not the expensive part. Computing scattering matrix elements is, and our method does not reduce that cost. The value of the reduced circuit is that it is portable. It gives closed-form temperature transfer, multi-pole $\sigma(\omega)$ and Hall response from one object, and it can be embedded in circuit simulators, which a k-grid solver cannot. If the temperature-extrapolation and frequency tests below fail, this argument fails with them.

Why not just fit a compact model or power law, as BSIM does? Empirical fits interpolate well. The evidence that they extrapolate poorly is physical: mobility exponents in Ti₂CO₂ shift with density and the dominant mechanism changes with temperature (Mandia et al.), so a power law fitted near room temperature should drift at 77 K and 500 K. The extrapolation experiment tests this directly.

Why not just train a neural surrogate on full-solve outputs? With few full solves it has no reason to respect conservation or to extrapolate. With many, it could match us inside its training range. We give it an equal budget and score it out of range.

Proposed Experiments

The first experiment measures accuracy against network size. For each material we sweep $K$ from 1 to 40 and plot mobility, Hall factor and $\gamma$ error relative to the full solve over the temperature–density grid. The decisive outcome is a $K \le 20$ network that recovers the Ti₂CO₂ Hall-factor range of 0.2–1.3 and the suspended-MoS₂ value near 1.43, while the diagonal ablation at the same $K$ visibly does not.

The second experiment tests temperature extrapolation. We reduce at 300 K only, apply the per-mechanism occupation scaling, and predict 77–500 K. We compare against the power law and the surrogate fitted on 250–350 K. The proposal fails if the network's error out of range is not clearly below the power law's.

The third experiment tests substrates. We use the same suspended-MoS₂ network topology and add mechanism-tagged remote-phonon edges for each dielectric, then check whether the substrate-induced mobility and Hall shifts follow without re-reducing the intrinsic part.

The fourth experiment tests frequency response and conservation. We compare $\sigma(\omega)$ from the reduced circuit against the full solve up to frequencies of a few times the inverse optical-phonon scattering time. We also confirm that the reduced operator keeps an exact zero mode, and that CRTA and the ablation do not.

A last experiment is a demonstration rather than a test. We insert exported subcircuits as device subgraphs in a terminal-explicit circuit graph and show temperature-swept channel conductance inside SPICE. No reported multi-corner circuit-learning result was found on Lacuna for comparison, so this part is presented as feasibility only.