The most common evidence that a crystal hosts "chiral" electrons is a drop in resistance when the magnetic field is aligned with the current. These are electrons that behave like massless particles with a fixed handedness, and the effect is called negative longitudinal magnetoresistance. The trouble is that a purely classical effect can fake it. In very clean, high-mobility crystals a strong field makes the material conduct far better along the field than across it, so the current squeezes into a narrow beam. Voltage probes away from that beam then read a smaller voltage, which looks like a resistance drop. This artifact is called current jetting.
Heavy-ion physics faced the same structural problem, where a quantum anomaly signal sits under classical backgrounds with the same footprint. It was attacked by machine-learning an observable that maximizes signal significance while forcing the background response to zero. We propose to carry that recipe into solid-state transport. The idea is to simulate a Weyl-semimetal sample with the anomaly switched on or off and with jetting switched on or off. From those simulations we learn a fixed combination of voltages from many contacts, field angles and field strengths that responds to the anomaly but is provably blind to jetting. We then test that combination on real crystals whose geometry and defects have been measured independently.
Core Claim
Optimal observables for the chiral magnetic effect from machine learning builds linear and bilinear combinations of flow harmonics and momentum weights. It trains them on 3×10⁶ simulated Au+Au events under four switches (chiral effect on/off, local charge conservation on/off). The resulting observable reached 18.7σ, with background contamination consistent with zero, against about 0.2 for the traditional correlators and a sensitivity gain of up to 90%.
The paper explicitly leaves open the solid-state case, where current jetting and ordinary magnetoresistance replace charge conservation and resonance decays as the confounders. Our claim is that the same factorial on/off design, applied to a transport simulator in which jetting arises naturally, yields a contact-voltage observable. That observable separates anomaly conduction from jetting more reliably than either the standard single-pair measurement or hand-designed geometry tests. The key new ingredient is the background-nulling constraint learned across held-out sample geometries. Without it, an optimizer will simply learn to exploit jetting, because jetting is the larger signal.
Research Program
The simulator supplies the method. Van Roosbroeck's equations with topological terms adds anomaly source terms and an intervalley relaxation time to standard semiconductor drift–diffusion equations, but solves them only in a linearized 1D thin film. We extend the steady-state version to 3D samples with a field-dependent anisotropic conductivity tensor and realistic contacts. This is what generates jetting, which 1D geometry cannot. The anomaly coefficient $C_A$ and the jetting-producing anisotropy are independent switches, giving a 2×2 design exactly parallel to the heavy-ion study.
The observable is
$$O = \sum_i w_i V_i(\theta,B) + \sum_{i,j} W_{ij}\, V_i V_j ,$$
where $V_i$ are voltages on contact pair $i$ at field angle $\theta$ and strength $B$, and $w, W$ are learned weights. Training maximizes the separation of $O$ between anomaly-on and anomaly-off samples in units of its noise. A penalty drives the mean of $O$ to zero in every anomaly-off configuration across a distribution of sample shapes, contact misalignments and mobility inhomogeneities. The noise model includes thermal and amplifier noise, so the significance is a real sensitivity estimate rather than a simulator artifact.
Multi-modal characterization enters as the constraint that makes the transfer to real crystals honest. Scanning electron microscopy fixes each crystal's actual geometry and contact placement, which are inputs to the simulator. Spatially resolved probing of the surface, in the spirit of the correlated AFM–NanoIR mapping of Makha et al., bounds surface damage and inhomogeneity. Surface damage matters especially for focused-ion-beam-cut devices such as the NbP channel of Rocchino et al., whose milling damage was never quantified. These measurements define the range of nuisance variation the observable must be robust over.
Baselines
The naive reference point is the standard four-probe longitudinal measurement of Zhang et al., who reported negative magnetoresistance of up to −30% at 1.8 K in TaAs, along with transverse magnetoresistance up to 80,000% and mobility of 1.8×10⁵ cm²/V·s. That last pair is exactly the regime where jetting is severe. We will run it on identical simulated and real samples and report its anomaly-off response. Our observable must show a much smaller anomaly-off response at equal or higher anomaly-on significance. No reported result exists on Lacuna for any quantitative jetting-contamination figure on this measurement.
The strongest physics-motivated competitor is a hand-designed multi-contact consistency test. Here voltages are compared on contacts along the sample's central axis versus its edges, and the anomaly is accepted only if they agree. This uses the same contacts and data as our method, with no learning. We will apply it under the same simulated nuisance distribution and the same real crystals.
The strongest statistical competitor is a full simulator-based likelihood fit that infers $C_A$ jointly with geometry and anisotropy nuisances. It gets the same simulator, data and compute as our training.
For the heavy-ion analog we cite the reported 18.7σ and background consistent with zero from Hirono et al. as the target behavior. We do not use it as a directly comparable number.
Two ablations isolate the key ingredient. The first removes the background-nulling penalty and keeps only significance. The second trains the nulling on a single geometry rather than a distribution. These test whether robustness across geometries, the limitation the heavy-ion work left open by using a single centrality, is what matters.
Why not just…?
Why not just collect more data with standard magnetoresistance? Averaging longer shrinks noise on a quantity whose bias is set by geometry. The heavy-ion analysis shows the same failure for traditional correlators, which stay at roughly 0.2 background contamination however many events are added. Our simulated anomaly-off runs will show the standard measurement's false signal is independent of averaging time.
Why not just do the full likelihood fit? This is a real threat. With a correct simulator it is the optimal analysis and may match our observable. Its weakness is misspecification, such as unmodeled inhomogeneity or contact damage, which biases the inferred $C_A$ directly. A fixed, mostly linear observable instead has a background response that can be checked on control samples. We will settle this by training both on one nuisance model and testing on deliberately different ones. If the likelihood fit is equally robust there, we will report it as the preferred method and recast the learned observable as a cheap, auditable screen.
Why not just shape samples into thin needles to suppress jetting? Geometry control helps and we will use it, but it cannot be verified from one voltage pair. It is also unavailable for small or irregular crystals and FIB devices. The geometry-distribution ablation tests whether learned weights add anything over well-shaped samples alone.
Decisive Checks
The central test is held-out geometry transfer. The observable is trained on one family of shapes and misalignments and evaluated on unseen ones. It must keep its anomaly-off response near zero while the standard measurement's false signal varies widely.
On real crystals of TaAs and NbP, the observable's anomaly signal must follow the expected dependence on the angle between field and current. It must also weaken with temperature as the negative magnetoresistance in Zhang et al. does. On a high-mobility control crystal without Weyl nodes, characterized by the same imaging suite, it must return a null result even though the standard measurement shows apparent negative magnetoresistance.
Because angle and field are scanned, all significances are reported after correcting for the number of angles and fields searched (a "look-elsewhere" correction). The project fails if the observable's real-crystal anomaly-off response exceeds its simulated noise floor. That outcome would mean the simulator misses a background, and finding which one is itself the result.
References
- Optimal Observables for the Chiral Magnetic Effect from Machine Learning
- Van Roosbroeck's equations with topological terms: The case of Weyl semimetals
- Insights into photovoltaic properties of ternary organic solar cells from phase diagrams
- Magnetoresistive-coupled transistor using the Weyl semimetal NbP
- Observation of the Chiral-Anomaly-Induced Negative Magnetoresistance in 3D Weyl Semimetal TaAs