Bayesian inference of high-purity germanium detector impurities based on capacitance measurements and machine-learning accelerated capacitance calculations

This paper presents a novel Bayesian inference method using a machine-learning surrogate model trained on GPU-accelerated simulations to accurately determine the spatially varying impurity density of high-purity germanium detectors from capacitance measurements, overcoming the limitations of traditional manufacturer data.

The Gist

The Big Picture: Tuning a Musical Instrument

Imagine you have a very expensive, high-precision musical instrument (a High-Purity Germanium detector). To make it play the right notes (detect rare physics events like dark matter), you need to know exactly how the wood inside is shaped and where the knots are. In physics terms, this "wood" is the germanium crystal, and the "knots" are impurities (tiny bits of other atoms mixed in).

The problem? The manufacturer gives you a rough guess of where the knots are, but it's like looking at a map drawn by someone who only walked a few steps. It's vague, and the "uncertainty" is huge. If you build your simulation (your mental model of the instrument) based on a bad map, your predictions will be wrong, and you might miss the music you're trying to hear.

The Solution: The "Capacitance Fingerprint"

The authors realized that the detector has a unique "fingerprint" called capacitance. Think of capacitance like the electrical "stiffness" of the detector.

• When you apply voltage (push the strings), the detector resists in a specific way.
• This resistance changes depending on how many impurities are inside and where they are located.

If you measure this "stiffness" at different voltages, you get a curve (a C-V curve). If you know the impurities, you can predict the curve. But the authors wanted to do the reverse: measure the curve and figure out the impurities.

The Problem: The "Slow Cooker" Simulation

Usually, to figure out the impurities, scientists have to run a computer simulation. They guess a set of impurities, run the math, see if the curve matches, guess again, and repeat.

• The Catch: Running this simulation is incredibly slow. It's like trying to bake a cake by checking the oven temperature every second, but the oven takes 10 minutes to heat up. Even with super-fast computers (GPUs), calculating one curve takes minutes. To find the perfect impurity map, you'd need to try millions of combinations. At that speed, it would take years.

The Innovation: The "Crystal Ball" (Machine Learning)

This is where the paper gets clever. The authors built a Deep Neural Network (DNN).

• The Training: They used their super-fast computers to calculate 60,000 different scenarios (different impurity maps) and recorded the resulting curves.
• The Learning: They fed this data into the AI. The AI learned the pattern: "Oh, if the impurities are high near the edge, the curve looks like this. If they are low in the middle, the curve looks like that."
• The Result: Once trained, the AI became a "Crystal Ball." Instead of taking 10 minutes to calculate a curve, the AI could predict it in microseconds (faster than you can blink).

Now, instead of waiting years, they could run a Bayesian Inference. Think of this as a super-smart detective who doesn't just guess; it weighs every possibility, updates its confidence as it gathers clues, and eventually narrows down the exact location of the impurities with a statistical guarantee.

The Discovery: The "Onion" Effect

When they applied this new method to their test detector (named "Super-Siegfried"), they found something surprising.

• The Old Belief: Scientists usually thought impurities changed only from the top of the crystal to the bottom (like layers in a cake).
• The New Reality: The data showed that impurities also change from the center to the edge (like layers in an onion). The edge of the detector had much fewer impurities than the manufacturer thought, and the distribution wasn't uniform.

Why Does This Matter?

1. Better Detective Work: In the hunt for dark matter or neutrinoless double-beta decay, scientists need to distinguish between a "signal" (a rare event) and "background noise." If your map of the detector is wrong, you might mistake noise for a signal or miss a real signal. Knowing the exact impurity map makes the detector's "ears" sharper.
2. Faster Design: This method isn't just for fixing old detectors. It can be used to design new detectors. You can ask the AI, "What impurity map would give us the best performance?" and get an answer instantly, speeding up the engineering process.

Summary

The paper is about building a super-fast AI assistant that learns the physics of germanium detectors. This assistant allows scientists to reverse-engineer the internal "fingerprint" of a detector just by measuring its electrical stiffness. They discovered that the detector's internal structure is more complex (radially dependent) than previously thought, which is a crucial step toward building better tools for exploring the secrets of the universe.

Technical Summary

1. Problem Statement

High-purity germanium (HPGe) detectors are critical for rare-event physics searches (e.g., neutrinoless double-beta decay, dark matter). Accurate simulation of these detectors requires precise knowledge of the electrically active impurity density distribution ($\zeta$) within the germanium crystal.

• Current Limitations: Manufacturer-provided impurity data (derived from Hall effect measurements) is typically limited to a few discrete locations (top/bottom of the crystal) and carries large uncertainties. Standard assumptions often model $\zeta$ as a simple linear or quadratic function along the crystal's vertical ($z$) axis, ignoring potential radial ($r$) variations.
• Consequences: Incorrect impurity models lead to inaccurate electric field calculations, which in turn distort charge carrier drift simulations and pulse shape predictions. This hinders the ability to distinguish signal from background events.
• Computational Bottleneck: While capacitance-voltage (C-V) measurements can theoretically constrain $\zeta$, performing full Bayesian inference on complex, multi-parameter impurity models is computationally prohibitive. Calculating a single C-V curve using standard field solvers (like SolidStateDetectors.jl) takes minutes, making the millions of evaluations required for Bayesian parameter exploration infeasible.

2. Methodology

The authors propose a novel workflow combining high-fidelity simulation, machine learning (ML), and Bayesian inference to solve the inverse problem of determining impurity distributions from C-V data.

A. Experimental Setup

• Detector: An $n$-type true-coaxial HPGe detector named "Super-Siegfried."
• Measurement: A C-V scan was performed over 60 bias voltages ($U_B$) ranging from 25 V to 2975 V.
• Data Extraction: Capacitance ($c_{12}$) was derived from induced charge pulses on the $n^+$ contact generated by a pulse generator injected into the $p^+$ contact. The analysis accounted for partial depletion effects at low voltages by fitting the pulse tails.

B. Simulation Framework (SolidStateDetectors.jl)

• The authors used SolidStateDetectors.jl (SSD), a Julia-based package, to simulate the electric potential and capacitance matrices.
• GPU Acceleration: Field calculations were performed on GPUs to speed up the generation of training data.
• Geometry Modeling: The simulation included the detector, a grounded holding structure, and an infrared shield ("hat"). The $n^+$ contact thickness ($d_{Li}$) was treated as a free parameter.

C. Machine Learning Surrogate Model

To bypass the computational cost of repeated field calculations during inference:

1. Data Generation: 60,000 quasi-random parameter sets were generated, varying the $n^+$ contact thickness ($d_{Li}$), the bias voltage ($U_B$), and 8 parameters defining the impurity density profile ($p_\zeta$).
2. Impurity Model: A flexible model $\zeta_{RZ}(r, z)$ was defined using cubic splines at the top and bottom of the detector, allowing for both radial and vertical dependence. A parameter transformation was applied to handle the wide dynamic range of impurity densities ($10^8$ to $10^{11}$ cm$^{-3}$).
3. Neural Network: A Deep Neural Network (DNN) was trained using Flux.jl to predict capacitance $c_{12}$ given the input parameters $(d_{Li}, p_\zeta, U_B)$.
   • Architecture: 7 layers (1 input, 5 hidden dense layers with GELU activation, 1 output).
   • Performance: The DNN predicts capacitance in microseconds ($\mu$s) with a relative error of $\approx 1.5\%$, compared to minutes for the full simulation.

D. Bayesian Inference

• Tool: BAT.jl was used to perform the Bayesian parameter estimation.
• Likelihood: Defined as the product of Gaussian distributions comparing measured C-V data against DNN-predicted curves, incorporating uncertainties from measurement, simulation, and the ML model.
• Models Tested:
  1. BZ (No Radial Dependence): Impurity varies only with $z$.
  2. BRZ (Radial Dependence): Impurity varies with both $r$ and $z$ using the full 8-parameter spline model.

3. Key Contributions

1. ML-Accelerated Bayesian Inference: The paper demonstrates a successful pipeline where a DNN surrogate replaces expensive physics simulations, enabling full Bayesian exploration of high-dimensional parameter spaces (10 dimensions) for detector characterization.
2. Discovery of Radial Impurity Dependence: The study proves that assuming a purely vertical impurity profile is insufficient. The data strongly suggests a radial dependence, with significantly lower impurity densities near the detector's outer mantle.
3. Open-Source Julia Ecosystem: The work showcases the integration of SolidStateDetectors.jl, Flux.jl, and BAT.jl for advanced detector physics, providing a reproducible, open-source framework.
4. Refined Impurity Model: Introduction of a flexible spline-based impurity model that allows for $n$-type to $p$-type transitions and radial variations, better reflecting physical realities like dopant diffusion or crystal growth artifacts.

4. Results

• BZ Model (Vertical Only): The Bayesian fit showed a bimodal posterior distribution (due to top/bottom symmetry) but failed to reproduce the measured C-V curve at low bias voltages ($<1000$ V). The residuals indicated that the detector depleted faster in reality than the model predicted, implying the actual impurity density at larger radii is lower than assumed.
• BRZ Model (Radial + Vertical):
  • The fit with radial dependence successfully described the entire C-V curve, with residuals centered around zero across all bias voltages.
  • Key Finding: The posterior distribution for impurities near the mantle (large $r$) favored very low densities, potentially indicating a $p$-type region or a significant depletion of $n$-type dopants near the surface. This is likely due to boron diffusion from the heavily doped $p^+$ contacts or radial segregation during crystal pulling.
  • The $n^+$ contact thickness ($d_{Li}$) was estimated to be approximately 3 mm, thicker than typical values, likely due to the detector's age and lithium drift growth.
• Impact on Pulse Shapes: Simulations using the newly derived impurity models ($\zeta_{BRZ}$) showed distinct differences in the electric field strength near the contacts compared to the manufacturer's model. This directly affects the simulated pulse shapes, highlighting the necessity of accurate impurity maps for pulse-shape analysis.

5. Significance

• Improved Detector Simulation: By accurately determining the impurity density distribution, simulations of electric fields and charge drift become more reliable. This is crucial for optimizing signal/background discrimination in rare-event searches.
• New Characterization Path: The method provides a non-destructive way to map impurity densities using standard C-V scans, complementing destructive Hall effect measurements.
• Future Applications: The framework is not limited to impurity fitting; it can be used for general detector optimization (e.g., finding optimal geometry parameters) and inferring impurity distributions from other voltage-dependent properties (e.g., depletion volume shape via Compton scanning).
• Methodological Advancement: It sets a precedent for using ML surrogates to make computationally expensive physics simulations tractable for rigorous statistical inference in experimental physics.