Quantum Computing News

QVAE Quantum Variational Autoencoders

Significant work has been conducted in applying quantum-AI to particle physics modelling as a result of a recent partnership between TRIUMF, Perimeter Institute for Theoretical Physics, and D-Wave Quantum Inc., specifically to address processing constraints for CERN’s Large Hadron Collider (LHC) upgrades. This groundbreaking study, which was published in npj Quantum Information, is the first instance of a quantum annealing device being used for the computationally costly particle shower simulation at the LHC.

You can also read QLASS European Use Glass & Light to Create Quantum Chips

The Challenge: Computational Bottlenecks at the LHC

LHC collides protons to measure and find particles like the Higgs Boson. After upgrading to “High-Luminosity LHC” (HL-LHC), collisions will increase tenfold. This improvement poses significant computational problems even though it will enable more accurate measurements and the detection of uncommon processes.

Designing future experiments, calibrating detectors, assessing data compliance with physical assumptions, and analysing existing experimental data all depend on the simulation of collision occurrences. Traditionally, first-principles particle simulation programs such as GEANT4 are used to carry out these simulations. Nevertheless, it takes about 1000 CPU seconds to simulate a single event using GEANT4, and during the HL-LHC phase, this computational intensity is expected to increase to millions of CPU-years per year, which is considered “financially and environmentally unsustainable”.

You can also read India Accelerates Quantum-Secure Satellite To Space

Simulating particle interactions with calorimeter components accounts for a large amount of this computing load. Calorimeters are detectors that use the showers that are created when particles pass through the active material of the detector to determine the particle energy. Accurately simulating these intricate particle showers is the most computationally demanding Monte Carlo (MC) simulation work, but it is essential for high-quality measurements.

A “CaloChallenge” was launched in 2022 to encourage progress in this field, offering datasets for groups to create and compare calorimeter simulations. It should be highlighted that this collaboration’s research team is the only one to far to tackle this problem from a full-scale quantum perspective.

The Quantum-AI Hybrid Solution: CaloQVAE and Calo4pQVAE

The team created CaloQVAE, a quantum-AI hybrid technique that was later enhanced to Calo4pQVAE, to overcome these issues. This method simulates high-energy particle-calorimeter interactions quickly and effectively by combining quantum annealing with current developments in generative models.

Fundamentally, Calo4pQVAE is conceived as a variational autoencoder (VAE) for which its prior is a limited Boltzmann machine (RBM). VAEs are a class of generative models for latent variables that approximate true log-likelihood by maximising an evidence lower bound (ELBO). The RBM improves the expressivity of the model by being a universal approximator of discrete distributions. Given particular incidence energy, the model is intended to produce artificial shower events.

Fully connected neural networks are used to model the encoder (qϕ(z|x,e)) and decoder (pθ(x|z,e)) components of the VAE, which are conditioned on incident particle energy. To account for the cylindrical geometry of the showers, Calo4pQVAE incorporates 3D convolutional layers and periodic boundary conditions. The approach assumes a Boltzmann distribution for the prior and employs a discrete binary latent space.

You can also read Quantum Crypto News: $8.6B Bitcoin Whale Movement Debates

The incorporation of D-Wave’s annealing quantum computing technology is a significant innovation. The researchers showed that they could use the D-Wave 2000Q annealer to create latent space samples in CaloQVAE. A masking function was developed to adapt the RBM to the QPU architecture (Chimaera graph topology), which is not entirely connected. In order to employ D-Wave’s more sophisticated Pegasus-structured Advantage quantum annealer for sampling, the two-partite graph of the RBM was swapped out for a four-partite graph for Calo4pQVAE.

Importantly, the team discovered that by unconventionally manipulating qubits, D-Wave’s annealing quantum computers could be used to generate simulations. They successfully “hijacked” a D-Wave quantum processor mechanism that typically maintains a steady ratio between a qubit’s bias and the weight connecting it to another. By fixing a subset of qubits (σz(k)), they may “condition” the processor and guarantee that these qubits stay in preset states during the annealing process. This implies that the system is capable of producing showers with particular desirable characteristics, such the energy of an impinging particle.

The flux bias parameters of the quantum annealer are used to accomplish this conditioning, which enables a flexible integration of classical RBM capabilities with the potential speedup and scalability of quantum annealing. Additionally, the study presents an adaptive technique for effectively calculating the quantum annealer’s effective inverse temperature a significant methodological breakthrough that could be advantageous for a variety of quantum machine learning applications.

You can also read Quantum Crypto News: $8.6B Bitcoin Whale Movement Debates

Performance and Benefits

The released findings demonstrate this quantum-AI hybrid approach’s encouraging performance on a number of metrics:

• Speed: Compared to generating samples via GPU, the raw Quantum Processing Unit (QPU) annealing time per sample is 20 µs, which is 20 times faster. The core annealing speed highlights the possibility for significantly outperforming classical methods with optimised engineering, even though the total quantum sampler rate (0.4 ms per sample) is marginally quicker than classical GPU approaches (~0.5 ms per sample). While the QA took about 0.1 seconds (assuming single QPU programming), the conventional approaches required about 1 second to generate 1024 samples.
• Accuracy and Fidelity: The CaloQVAE model produces synthetic data that accurately reflects significant patterns found in the real data. Its accuracy metrics for particle categorization (e.g., e+ vs. π+) are extremely similar to those of other techniques, including CaloGAN. The results of GEANT4 data and the generative models closely match, according to qualitative examination of shower form variables, suggesting that the generative models accurately capture key characteristics and associations. The quantum device’s sample quality is comparable to that of contemporary Monte Carlo techniques. Both classical (DVAE) and quantum (QVAE) methods were able to replicate the features seen in real GEANT4 data for the energy conditioning of the model. Based on Fréchet physics distance (FPD) and Kernel physics distance (KPD) measures, this framework outperforms almost half of the models that were evaluated in the CaloChallenge.
• Computational Efficiency/Energy Consumption: The energy consumption is a defining characteristic. D-Wave quantum processors use the same amount of energy regardless of job size, unlike classical GPUs. This suggests that QPUs could develop without needing more processing power, making high-demand simulations more feasible in the future.

You can also read Quantum Proof-of-Work QPoW Simulator By BTQ Technologies

Collaborating Institutions and Future Implications

TRIUMF, Perimeter Institute for Theoretical Physics, and D-Wave Quantum Inc. collaborated on this important study. The University of Virginia, the University of British Columbia, and the National Research Council of Canada (NRC) all made further contributions.

In order to improve their models’ speed and accuracy, the team intends to test them on fresh incoming data. In order to improve simulation quality, they plan to upgrade to D-Wave’s most recent quantum annealer (Advantage2_prototype2.4), which offers more couplers per qubit and lower noise, investigate various RBM configurations, and improve the decoder module.

If scalable, this approach can be used to generate synthetic data for industries like manufacturing, healthcare, and finance, among other areas beyond particle physics. Larger-scale quantum-coherent simulations as priors in deep generative models are anticipated as a result of the authors’ belief that annealing quantum computing will become a crucial component of simulation generation. A promising use of quantum computing to address unresolved basic physics research issues is demonstrated by this work.

You can also read Power Of Qutrit Entanglement: Beyond Qubits In Quantum Tech