BMW, Zapata and MIT Test Quantum-Inspired Generative AI in Production

Zapata’s Generator-Enhanced Optimization (GEO) outperformed traditional problem solvers in 71 percent of cases

Scheme for Generator-Enhanced Optimization (GEO) strategy. (Source: Zapata)

Scheme for Generator-Enhanced Optimization (GEO) strategy. (Source: Zapata)

Quantum technologies are beginning to reshape the landscape of industrial automation for the automotive and other sectors. A recent collaborative effort between Zapata Computing, BMW and MIT’s Center for Quantum Engineering (CQE) showcased the potential of quantum technologies to boost the efficiency and performance of automotive production lines.

The partnership yielded groundbreaking research in using quantum-inspired generative AI techniques to optimize vehicle production. This blend of AI and quantum computing could mark the dawn of a new era in automotive manufacturing, highlighting the potential of quantum technologies to revamp complex industrial processes.

The collaboration addressed one of the most crucial challenges in the automotive industry—optimizing vehicle production schedules across multiple manufacturing plants. At its core, this challenge is a convoluted puzzle with countless configurations and constraints, ranging from fluctuating production rates, discrete shift schedules to the prevention of buffer overflows and shortages during manufacturing processes.

GEO Versus Solvers 

Zapata Computing’s Generator-Enhanced Optimization (GEO) technique was put to the test, conducting approximately one million optimization runs in a series of simulations. These trials were run on Zapata’s Orquestra software platform, comparing the efficiency and performance of GEO against various algorithms and problem configurations. The study’s results demonstrated that GEO outperformed even state-of-the-art solvers in minimizing assembly line idle time while meeting monthly vehicle production targets.

“We ran roughly a million optimization runs cycling through dozens of various algorithms, problem configurations and optimizer solutions to benchmark their performance against each other,” said Yudong Cao, CTO and co-founder at Zapata Computing. “GEO uses quantum or quantum-inspired generative machine learning models to learn from and improve upon the results generated by classical solvers.”

The endeavor not only provided BMW with potential production solutions but also served as a salient example of quantum computing’s application in tackling real-world industrial problems. Dr. William D. Oliver, Director of The Center for Quantum Engineering, hailed the project as a prime example of the center’s mission to link academic rigor with industry partners to solve practical challenges.

The GEO technique leverages quantum-inspired generative machine learning models, which draw from principles of quantum mechanics like entanglement and superposition but don’t require use of a quantum computer. The technique is used to learn from and enhance the outcomes produced by classical solvers. GEO shows how quantum technologies can be harnessed to enhance computational outcomes despite quantum computers still being in the early stages of R&D.

For BMW, the application of quantum-inspired generative AI techniques is a promising stride towards revolutionizing their manufacturing operations. The successful benchmarking of GEO against other production planning techniques aligns with BMW’s drive for innovation and operational efficiency.

Conflicting Constraints

The significance of this research extends beyond the automotive industry. As we inch closer to a quantum future, collaborations like these lay the groundwork for broader applications of quantum technologies in fields like medicine, materials science, financial forecasting, energy, and climate science.

The study could also help enhance the efficiency of production lines using quantum-inspired techniques. Cao explained that the optimization challenge lies in finding the best approach under conflicting constraints in complex production settings.

“When running a production line, you want to decide what the production rate should be, how fast each station should go,” said Cao, “But in reality, workers cannot go forever, they have their shifts, they have their breaks. You need to arrange these steps in a way that minimizes inventory and space. That’s part of the optimization problem.”

Traditional methods for solving such problems—such as linear arrangements—have limitations, especially when there are multiple variables at play. This optimization problem is like a difficult math problem where all possible solutions must be exhaustively explored, which is often not feasible in a large-scale business context. One caveat to some current aspects of the optimization problem is the rising deployment of robotics in manufacturing, which could yield more predictability compared to human workers.

Different techniques have been explored to solve these complex optimization problems. “In this project with BMW, we actually used quite a few techniques—simulated annealing, genetic algorithms, Monte Carlo algorithms, etc,” said Cao. “But what we showed in this project, is that you can boost the performance of these optimizers even further with AI. That’s the gist of what GEO is about—using AI combined with what people have come up with so far for solving optimization problems.”

One of the study’s goals was to minimize inventory. “We benchmarked GEO against classical algorithms for many variants of the problem,” said Cao. “We found that in 71 percent of cases, GEO either ties or outperforms state-of-the-art solvers.”

Cao also indicated that a quantum circuit could easily replace a standard one, explaining, “The way the GEO algorithm is set up, we can pretty much just swap out the tensor network construction and swap in the quantum circuit.”

Benchmarking and Scale-up

Cao noted that further benchmarking is needed to scale up the problem-solving capacity, by considering infrastructure aspects such as data ingestion and the frequency of algorithm execution. Additionally, he mentioned the potential of switching to a quantum solution, “just by swapping out the tensor networks with quantum circuits,” which would involve changing just one line of code.

The role of Orchestra has been critical in facilitating benchmarking, having run millions of instances of the algorithm. The next step in the study is to run more benchmarks on larger problems, with a goal to use this technology for production use cases.

Cao elaborated on the ongoing work to use quantum computing to solve complex optimization problems. Zapata researchers introduced a “synergistic framework” where “an algorithm acts like a relay race between a classical computer and a quantum computer.” The classical computer is utilized to solve as much of the problem as possible before handing over the more complex elements to the quantum computer.

Their theory outlines how tensor networks, a quantum-inspired data structure, can be mapped to a quantum device for use on a specific application.

“The quantum-inspired approach is a lot easier to work with because you can just run it on CPUs and GPUs, it’s just linear algebra under the hood,” said Cao. “We don’t have to submit a job, pick a hardware backend, look at the parameters. When using a quantum device there are extra steps that need to be taken in order to know that what you’re trying to do on the device is what the device is actually doing.”

However, Cao anticipated that using quantum computers could potentially lead to performance gains and efficiency, noting that quantum computers might be able to discover solutions that would take longer using a classical computer. At the same time, more research will be needed to define what specific problems are amenable to faster solving with quantum computers.

In terms of the potential industry impact, the GEO technique isn’t limited to automotive. Cao underscored that this hybrid scheme of classical and quantum-inspired techniques could tackle “the same kinds of operations research problems that people in Fortune 100 companies have been grappling with for decades.” The approach, as demonstrated in the BMW case study, allows for quantum technologies to enter the arena of complex operational challenges, acting as a competitor against other commercial solvers.