GeoRead: Geometry-Aware Transformer for Continuous AC-LGAD Readout Reconstruction

arXiv:2611.08694 · hep-ex, physics.data-an, physics.ins-det · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.

Kai: We're looking at the paper: "GeoRead: Geometry-Aware Transformer for Continuous AC-LGAD Readout Reconstruction".

Mira: Here's the gist of the paper. AC-coupled Low-Gain Avalanche Diodes (AC-LGADs) are promising silicon sensor technologies for high-precision four-dimensional tracking Position reconstruction in AC-LGADs relies on correlations among waveforms induced on multiple readout pads, where analytically solving the signal propagation becomes difficult, motivating data-based machine learning reconstruction GeoRead introduces a spatial-temporal Transformer enabling multi-event continuous position prediction and uncertainty estimation.

Kai: Problem Context Precision timing and spatial resolution are indispensable for next-generation experiments such as the HL-LHC

1, 2: , future e+e− colliders, and the EIC, where 4D tracking suppresses pile-up, enables time-of-flight particle identification, and reconstructs dense vertices.

Mira: The correlations among pad signals depend on hit position, readout geometry, and signal propagation through the resistive layer, making reconstruction particularly challenging.

Kai: Existing machine learning approaches generally assume a fixed readout pad layout and predict one hit from an isolated, temporally aligned waveform window.

Title and authors: Kai: Now moving into the specifics of what they actually built in "GeoRead: Geometry-Aware Transformer for Continuous AC-LGAD Readout Reconstruction." The authors are Lixing Wang, Leena Diehl, Spencer Ellis, Marcos Fernández García, Gabriele Giacomini, Vagelis Gkougkousis, Ben Kilminster, Jingyu Luo, Anna Macchiolo, Yajna Mahabeer.

Mira: They are coming from different areas—quantum hardware and condensed matter theory—which is good because it gives you a broad perspective on what they’re trying to achieve.

Lev: From my side, I focus on how this would run on real hardware, so I'm interested in the practical aspects of the AC-LGAD setup and how much overhead this ML model adds to the system.

Kai: I'm looking at what they built and measured, so they used four channel transient current technique data from a five hundred micrometer pitch AC-LGAD for their performance assessment <ref:2611.08694#pg3>.

Mira: And that’s where the context comes in because those sensors are promising for high precision four dimensional tracking, which is needed for experiments like the HL-LHC and future e plus e minus colliders.

Lev: So they're aiming at those demanding physics environments where timing and spatial resolution are absolutely indispensable to suppress pile up and reconstruct dense vertices.

Kai: The paper mentions that LGADs can meet timing requirements with performance around twenty to fifty picoseconds, which is the value they’re trying to achieve with these sensors.

Mira: But the typical ones have small pad sizes, like one by one millimeter squared, which limits spatial resolution pretty significantly on their own.

Lev: This is why they’re pushing for this kind of enhancement; they want to overcome those inherent size limitations in the silicon sensor design itself.

Kai: GeoRead aims to address that by using correlations among waveforms induced on multiple readout pads where analytically solving signal propagation becomes difficult, which motivates them to use data based machine learning reconstruction.

The paper's summary: Mira: So the paper summarizes how they tackled that correlation problem by moving away from fixed layout assumptions and instead focusing on a spatial-temporal approach.

Kai: They found that the key difficulty is that the signal propagation through the resistive layer depends not just on where you land, but also where you are.

Lev: That dependency on position makes it hard to model analytically, so ML is motivated to learn these complex correlations from waveforms instead of trying to solve the physics equations perfectly.

Mira: They use a Transformer architecture that incorporates rotary position encoding for temporal order and inject pairwise pad geometry into the attention in a way conceptually related to the Particle Transformer twenty-four.

Kai: The architecture has this geometry aware multi head attention, which applies rotary position encoding to queries and keys and uses a learned attention bias derived from relative pad geometry.

Mira: They also have a per time step AdaLNSN that aggregates hidden states across pads only at the same time index and combines this with an embedding of the bias voltage to determine sample dependent modulation.

Lev: That aggregation step is where you can see how they are combining spatial and temporal info at discrete moments in time, which is a crucial part of the methodology.

Kai: They also have a pad weighting position head that outputs a point prediction as a weighted average of known pad coordinates, allowing for out of hull prediction by permitting negative weights.

Mira: And finally, they have that separate uncertainty head predicting a positive definite two by two covariance matrix using the two dimensional Gaussian negative log likelihood.

Lev: So the model is designed to be comprehensive, handling geometry and time explicitly and then providing an estimate along with a measure of uncertainty for every time step.

The paper's improvements: Kai: The paper suggests three major improvements to their approach that go beyond what they initially proposed. First, the geometry aware attention in GeoRead decouples absolute sensor geometry from the reconstruction task by incorporating relative pad geometry directly into attention via a head dependent bias.

Mira: That’s a big conceptual win because it means they can generalize across sensor geometries, which is good because it lets you apply this reconstruction method to different detector designs without retraining everything for every new layout.

Lev: So instead of building a model specific to one fixed layout, the model learns the geometric relationships generally.

Kai: Second, the spatial temporal formulation enables native multi event continuous AC-LGAD position reconstruction, which allows the model to process temporally unaligned multi channel signals and support temporally overlapping signals without predefined event windows.

Mira: That means they can handle those messy data streams that are common in real experiments where things aren't perfectly aligned from the start.

Lev: It moves them away from having to wait for perfect alignment, which is a major operational improvement for continuous data acquisition.

Kai: Third, the training pipeline based on time jitter and event mixing relaxes assumptions by using random temporal shifts and additive superposition of independent event waveforms.

Mira: That allows the model to handle data synthesized from single event measurements more effectively because it teaches it how to deal with imperfect input data.

Lev: So they are essentially building in robustness against the fact that the training data might not perfectly represent clean, isolated events.

Conclusion: Kai: To wrap up, the main contributions of "GeoRead: Geometry-Aware Transformer for Continuous AC-LGAD Readout Reconstruction" are its geometry aware formulation that decouples absolute sensor geometry from the reconstruction task.

Lev: Plus, they provide predictive uncertainty, which is a measure of reconstruction quality, allowing the model to output both a hit position estimate and a predictive covariance for each time step.

Kai: They show that by incorporating relative pad geometry directly into the attention via a head dependent bias, you can decouple absolute sensor geometry from the reconstruction task.

Mira: And they introduce this spatial temporal formulation for native multi event continuous AC-LGAD position reconstruction.

Lev: And they provide predictive uncertainty, which is a measure of reconstruction quality, allowing the model to output both a hit position estimate and a predictive covariance for each time step.

Kai: They also show that training with time jitter and event mixing relaxes peak alignment and single event assumptions by using random temporal shifts and additive superposition of independent event waveforms.

Lev: So, it’s a big step forward for continuous position updates on synthetic multi event waveforms formed by superimposing temporally shifted TCT signals.

Kai: And we have to keep in mind their limitations, including the assumption of linear waveform superposition for multi event evaluation and the lack of explicit accounting for channel dependent response variations in the geometry aware attention.

Mira: And they also flag that computational cost and inference latency for continuous reconstruction still need to be thoroughly evaluated.

Lev: That’s what GeoRead: Geometry-Aware Transformer for Continuous AC-LGAD Readout Reconstruction is, it lays out a spatial temporal formulation for native multi event continuous AC-LGAD position reconstruction.

Mira: And it introduces this spatial temporal formulation for native multi event continuous AC-LGAD position reconstruction.

Lev: And they provide predictive uncertainty, which is a measure of reconstruction quality, allowing the model to output both a hit position estimate and a predictive covariance for each time step.

Brown University · University of Zurich · Instituto de Físia de Cantabria (CSIC-UC) · Brookhaven National Laboratory

hep-ex, physics.data-an, physics.ins-det

Submitted: 2026-11-05

Updated: 2026-11-10

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: AC-coupled Low-Gain Avalanche Diodes (AC-LGADs) are promising silicon sensor technologies for high-precision four-dimensional tracking Position reconstruction in AC-LGADs relies on correlations among…

Key concepts

Geometry-aware Multi-head Attention (Geo-MHA)
This mechanism integrates the relative positions of readout pads directly into the attention calculation. It uses rotary position encoding combined with a learned bias derived from pad geometry to allow the model to generalize across different sensor layouts, decoupling reconstruction accuracy from absolute sensor placement.
Spatial-temporal Formulation
The architecture is designed for continuous tracking, meaning it can process signals that are not perfectly aligned in time or space. It uses a Transformer structure combined with a per-time-step mechanism to predict hit positions across multiple events simultaneously, handling temporally overlapping signals without needing predefined event boundaries.
Time Jitter and Event Mixing
These data augmentation techniques simulate real-world signal misalignment. Time jitter randomly selects windows around event peaks, while event mixing superimposes shifted waveforms. This training approach forces the model to learn robust features that work even when input events are temporally misaligned or overlapping.
Predictive Uncertainty Head
This component predicts a positive-definite 2x2 covariance matrix for each time step. This output provides a measure of reconstruction quality, giving the user not just a position estimate but also an associated confidence level, which is crucial for reliable downstream physics analysis.

Terminology

Summary

AC-coupled Low-Gain Avalanche Diodes (AC-LGADs) are promising silicon sensor technologies for high-precision four-dimensional tracking Position reconstruction in AC-LGADs relies on correlations among waveforms induced on multiple readout pads, where analytically solving the signal propagation becomes difficult, motivating data-based machine learning reconstruction GeoRead introduces a spatial-temporal Transformer enabling multi-event continuous position prediction and uncertainty estimation. p. 0

Problem Context

Precision timing and spatial resolution are indispensable for next-generation experiments such as the HL-LHC [1, 2], future e+e− colliders, and the EIC, where 4D tracking suppresses pile-up, enables time-of-flight particle identification, and reconstructs dense vertices. The correlations among pad signals depend on hit position, readout geometry, and signal propagation through the resistive layer, making reconstruction particularly challenging. p. 2 Existing machine learning approaches generally assume a fixed readout pad layout and predict one hit from an isolated, temporally aligned waveform window. p. 0

Model Architecture Details

GeoRead incorporates relative pad positions into attention and produces time-resolved hit position estimates with predictive uncertainties. The architecture builds on the Transformer [26], uses rotary position encoding to represent temporal order [25], and injects pairwise pad geometry into attention in a way conceptually related to the Particle Transformer [24]. p. 3 Key architectural components include:

  1. Geometry-aware multi-head attention (Geo-MHA), which applies rotary position encoding to queries and keys and uses a learned attention bias derived from relative pad geometry. p. 5

  2. A per-time-step AdaLNSN, which aggregates hidden states across pads only at the same time index and combines this with an embedding of the bias voltage to determine sample-dependent modulation. p. 5

  3. A pad-weighting position head that outputs a point prediction as a weighted average of known pad coordinates, allowing for out-of-hull prediction by permitting negative weights. p. 6

  4. A separate uncertainty head that predicts a positive-definite 2x2 covariance matrix using the two-dimensional Gaussian negative log-likelihood. p. 6

Training and Data Augmentation

To relax the assumptions of peak alignment and isolated events, the model is trained using two data augmentations:

  1. Time Jitter: This simulates misaligned events by choosing a random window around each event’s peak, which is done by selecting a smaller window from a parent window. p. 4

  2. Event Mixing: This constructs multi-event waveforms by superimposing temporally shifted event waveforms, where subsequent events are added at sampled temporal offsets. p. 4

The training pipeline is further enhanced by evaluating the model across all available voltage points and training jointly on all available voltage points and respectively evaluating on each. p. 6

Performance Metrics

Using four-channel transient-current-technique (TCT) data of a 500 μm-pitch AC-LGAD, GeoRead achieves median position errors of 5.8 μm within the convex hull of the readout-pad centers and 8. p. 02 μm over the full scanned region at 190 V. p. 0 The model maintains comparable accuracy under temporal shifts and demonstrates time-resolved reconstruction of superposed events, without predefined per-event window. p. 0 The performance is assessed by comparing the model with its full-waveform variant as baseline. p. 3

Conclusion and Limitations

The main contributions are a geometry-aware formulation that decouples absolute sensor geometry from the reconstruction task, a spatial-temporal formulation for native multi-event continuous AC-LGAD position reconstruction, a training pipeline based on time jitter and event mixing that relaxes peak-alignment and single-event assumptions, and predictive uncertainty. The model demonstrates continuous position updates on synthetic multi-event waveforms formed by superimposing temporally shifted TCT signals. p. 4 Several limitations remain, including the assumption of linear waveform superposition for multi-event evaluation and the lack of explicit accounting for channel-dependent response variations in the geometry-aware attention. p. 3 The computational cost and inference latency of continuous reconstruction also remain to be evaluated. p. 6 The model also demonstrates continuous position updates on synthetic multi-event waveforms formed by superimposing temporally shifted TCT signals (Figure 2). p. 4

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Geometry-aware attention in GeoRead decouples absolute sensor geometry from reconstruction task, meaning it allows for generalization across sensor geometries by incorporating relative pad geometry directly into attention via a head-dependent bias.

  2. Spatial-temporal formulation enables native multi-event continuous AC-LGAD position reconstruction, allowing the model to process temporally unaligned multi-channel signals and support temporally overlapping signals without predefined event windows.

  3. Training pipeline based on time jitter and event mixing relaxes assumptions by using random temporal shifts and additive superposition of independent event waveforms, enabling the model to handle data synthesized from single-event measurements more effectively.

  4. Predictive uncertainty provides a measure of reconstruction quality, allowing the model to output both a hit-position estimate and a predictive covariance for each time step, which can be used for assessing reliability in downstream reconstruction.

  5. The model can perform continuous position updates on synthetic multi-event waveforms formed by superimposing temporally shifted TCT signals, supporting the reconstruction of successive hits without predefined event windows.

Sources

Related papers