Quantum Computing and AI Combine to Accelerate Complex Problem Solving

Researchers are continually seeking methods to enhance the practicality of quantum computing for complex optimisation challenges. Ying-Wei Tseng, Yu-Ting Kao, and Yeong-Jar Chang from the Industrial Technology Research Institute, alongside Jia-Han Ou and Wen-Zhi Zhang from National Pingtung University, demonstrate a hybrid quantum-classical approach to achieving sub-exponential speedup in gate-based quantum computing for Quadratic Unconstrained Binary Optimisation problems. Their work addresses the limitations of current gate-based quantum algorithms, which typically offer only quadratic speedup, and presents a method integrating Simulated Annealing to improve performance on large-scale problems. The team illustrate the significance of this advancement through a case study involving enzyme fermentation, where optimising 625 binary parameters to maximise active ingredient production is formulated as a challenging QUBO problem, proving the potential for industrial application.

Combining artificial intelligence and quantum computing for optimised experimental design

Scientists are increasingly focused on optimizing complex systems through combinatorial approaches, particularly in biotechnology and enzyme fermentation. Traditional methodologies like Design of Experiments (DoE) and response surface methodology (RSM) require extensive physical trials, incurring substantial costs and extending timelines.
The advent of Artificial Intelligence (AI) offers a streamlined approach to experimental design with fewer trials, while quantum-inspired computing enhances computational power for efficient searches within large combinatorial spaces. Classical Simulated Annealing (SA), introduced by Kirkpatrick et al., remains a widely used baseline method for high-dimensional optimization.

Glover and Kochenberger later formalized the QUBO framework as a universal modeling tool for NP-hard problems. In enzyme fermentation, differing enzyme formulations contain varying Active Ingredient (AIN) values, and the exponential growth in possible formulations makes improving AIN levels through physical trials increasingly challenging.

This study applies a QUBO model to estimate AIN in real-world fermentation processes. Due to the nonlinear nature of biological systems, quadratic models inevitably introduce distortions. Using the traditional Mean Square Error (MSE) cost function, experiments with 18 and 405 trials yielded errors of 19.99% and 15.51%, respectively, showing limited improvement despite increased data.

Conversely, employing the proposed Contour-Aware Cost Function reduced errors to 8.95% and 0.78% for the same experiments. The 0.78% error demonstrates the QUBO model’s ability to predict AIN with high precision for top-performing formulations, which is particularly meaningful for practical optimization.

Recent advances highlight the value of applying AI and machine learning to biological optimization, including enzyme production and bio-process modeling, and deep learning applications in genomics and biomedicine. Simulation and analysis are challenging because quantum states require 2-scale matrix and vector operations, causing design complexity and analytical errors to grow exponentially.

Circuit design also remains difficult, as leveraging entanglement and superposition offers parallelism but demands increasingly sophisticated architectures. This study proposes a new method to optimize the enzyme fermentation process, consisting of converting fermentation factors to binary variables, using machine learning with known experimental data to optimize and generate QUBO coefficients, including 484 quadratic terms, 22 linear terms, and one constant term, totaling 507 coefficients, utilizing QUBO software simulation to identify better fermentation formulas, and conducting targeted experiments to further optimize the formula.

Assuming two possible states, the cost function is approximately Gaussian distributed. Scenario 1 assumes Manufacturer A guesses a state with a cost of −σ, and SA requires 27 iterations to achieve a 99% chance of a result better than Manufacturer A. Scenario 2 assumes Manufacturer B uses a heuristic to find a solution with a cost of −2σ, requiring 200 iterations of SA for a 99% chance of achieving a better result.

The effectiveness of SA is related to the data distribution and the ability of heuristic algorithms, with simulation time increasing exponentially for each improvement of σ. An improved data augmentation and elimination method, termed “Walking in the Snow in Search of Plum Blossoms,” is proposed to enhance the active ingredients (AIN) in enzyme fermentation. The research team encoded 625 binary parameters representing variables such as temperature, stirring frequency, pH value, tryptophan concentration, and rice flour content, thereby defining the formulation space for enzyme production.

Minimizing the QUBO cost function directly corresponds to maximizing the AIN, framing the optimization problem as a search for the optimal binary configuration. Initial experiments, consisting of 600 physical trials, failed to improve AIN levels, highlighting the limitations of traditional optimization methods in this complex, multi-variable system.

To address this, the work implemented a hybrid approach integrating Artificial Intelligence (AI) with the QUBO model, aiming to accelerate the identification of superior enzyme formulations. The team conducted comparative experiments, initially performing 18 trials and subsequently expanding to 405 trials, using both traditional methods based on Mean Square Error (MSE) and the newly proposed Contour-Aware Cost Function.

Traditional methods yielded errors of 19.99% with 18 trials and 15.51% with 405 trials, demonstrating marginal improvement despite increased experimental effort. Notably, the Contour-Aware Cost Function significantly reduced errors to 8.95% with 18 trials and achieved an exceptional 0.78% error with 405 trials.

This substantial reduction in error indicates the QUBO model’s capacity to accurately predict AIN levels for high-performing formulations, representing an 18.7% increase from an initial AIN of 8481 to 10068. The methodology’s success stems from its ability to navigate the nonlinearities inherent in biological systems, offering a software-driven alternative to resource-intensive physical experimentation.

Hybrid optimisation enhances active ingredient yield via contour-aware quadratic modelling

Active Ingredient (AIN) levels in enzyme fermentation were improved from 8481 to 10068 through a novel hybrid optimization method, representing an 18.7% increase. Traditional modeling using the Mean Square Error (MSE) cost function resulted in errors of 19.99% after 18 experimental trials and only decreased to 15.51% with 405 trials.

Implementing a Contour-Aware Cost Function significantly reduced these errors to 8.95% after 18 trials and further to 0.78% after 405 trials. This 0.78% error rate demonstrates a high degree of precision in predicting AIN values for optimal enzyme formulations. The study utilized a QUBO model to estimate AIN in real-world fermentation processes, addressing the challenges posed by the nonlinear nature of biological systems.

The proposed method offers a software-driven approach to reduce experimental cycles and development timelines, particularly valuable given that each fermentation experiment requires over one hour to complete. The study establishes that Simulated Annealing, when integrated with quantum computation, can significantly reduce the computational burden compared to exhaustive search methods.

Specifically, the hybrid method achieves a complexity of approximately O(n2), a substantial improvement over the O(2n) complexity of brute-force approaches. Simulations indicate that the number of samples needed to achieve comparable solution quality remains relatively stable even as the problem size increases from 100 to 1000 bits, suggesting that computation time does not scale exponentially with problem size.

The effectiveness of the method is linked to the distribution of the cost function and the quality of heuristic algorithms employed. The authors acknowledge that achieving substantial improvements in solution quality still requires considerable computational resources, with even a seven-fold increase in runtime yielding only a marginal improvement of approximately one standard deviation.

Future work will focus on conducting more targeted experiments based on the simulation results to further refine the fermentation formula. The research highlights the potential of hybrid quantum-inspired algorithms for tackling real-world optimization challenges, although limitations remain in terms of scaling and achieving significant gains with increased computational effort.