Quantum machine learning (QML) offers a promising pathway for early adoption in industrial settings. For QML solutions to be effectively implemented in production, dedicated access to quantum hardware is essential. The quantum hardware ecosystem is rapidly evolving, with most of the infrastructure hosted by cloud providers. However, this reliance on cloud services presents challenges, including latency, security concerns, and high costs which can impede real-time applications. Is there a way to circumvent these limitations?
December 3, 2025
© iStock/peterschreiber media
As the world shifts towards renewable energy, consumers are increasingly becoming "prosumers"—individuals who are both consumers and producers—making it more challenging to maintain grid stability. Major energy utility providers such as E.ON must contend with the volatility arising from weather-dependent generation and fluctuating demand, which calls for innovative approaches to monitor and forecast consumption and production in real time.
Quantum machine learning (QML) holds promise: although definitive, large-scale advantages have yet to be demonstrated, early indications—such as HSBC and IBM’s quantum-enhanced financial trading pilot application [1]—point to its potential. Furthermore, there are compelling reasons to anticipate that quantum learning methods could enhance performance under specific conditions.
The core challenge, however, is operational: how do you deploy a QC-powered solution in a decentralized, secure way that can deliver real-time predictions across the grid?
A diverse “zoo” of quantum hardware is currently emerging: although much of it remains in research laboratories, a growing portion is now commercially available. Major cloud providers—IBM (IBM Cloud), Amazon (AWS Braket), and Microsoft (Azure Quantum)—are developing quantum capabilities and hosting access to quantum hardware on their platforms. Additionally, companies like IQM and AQT are integrating their systems into High Performance Computing (HPC) infrastructures, such as the Leibniz Supercomputing Centre.
While the widespread availability of quantum hardware allows virtually anyone to develop prototypes and test quantum algorithms, it also presents several challenges: lengthy job queues and resulting latencies that are incompatible with real-time applications; security risks associated with transmitting sensitive data (such as information about electrical infrastructure) to third-party cloud providers; and high operational costs that escalate rapidly with frequent access to quantum hardware.
Despite rapid technological advancements and an expanding ecosystem of hardware vendors, acquiring a dedicated on-premise quantum computer remains unrealistic for most companies in the foreseeable future. Consequently, it is logical to explore alternative deployment models for quantum solutions.
To avoid latency and security issues in productive operation, inference on classical hardware is recommended. Simulating quantum circuits on classical machines is known to be complex and makes real-time applications unrealistic. So, can the input-output relationship of a trained quantum model be approximated directly instead of simulating it (which is more complex)?
This concept, referred to as classical surrogates [2], enables us to achieve this goal. It is based on the observation that many parameterized quantum circuits display periodic behavior in their outputs. This periodicity can be effectively captured using a Fourier series, which serves as a functional approximation of the model's response.
The evaluation of a Fourier series can be conducted efficiently with minimal computing resources, without the need for a quantum processing unit (QPU) or a graphics processing unit (GPU). As a result, surrogate inference can be performed quickly on standard CPUs, including those in resource-constrained edge devices. This positions classical surrogates as a practical solution for bridging the gap between training on quantum hardware and integrating into existing traditional infrastructures.
Returning to the previously introduced use case, one could train a quantum model in a development environment e. g. to predict the net hourly electrical power production and consumption of a power plant, providing critical information for power grid stability. If the quantum model shows performance advantages, a classical surrogate can then be developed for direct deployment on-site at the power plant. However, a key question remains: how easy is it to create classical surrogates?
In a joint research project between E.ON and Fraunhofer Institute for Cognitive Systems IKS, the objective was to develop an end-to-end surrogation pipeline for a quantum model. This model was trained to forecast hourly net electricity production and consumption in E.ON's combined heat and power plants. The findings highlighted a significant limitation [3] of a previously proposed method for classical surrogates [2] which restricts its applicability on an industry-relevant scale.
During the construction of Fourier series of the original method [2], a large matrix must be stored. This matrix was so extensive that even a quantum circuit with a 13-qubit width and 2-layer depth required an High Performance Computing (HPS) system. Furthermore, the space requirements scaled exponentially, rendering this method unsustainable for industry-relevant application scales that could involve hundreds of qubits.
The most significant outcome of the joint project was the development of an alternative method (here referred to as Surrogation v2.0) which substantially reduces computational resource requirements by eliminating redundancies that, while providing exact guarantees, are irrelevant for practical applications. We created a proof of concept on a standard laptop using a 9-qubit, 2-layer model. This new surrogate required only 0.3 x 10-9 percent of the information previously needed (approximately 10 TB was reduced to just 16 GB of RAM). As a result, we successfully surrogated a quantum model for power usage prediction with negligible loss in accuracy. Moreover, our method exhibits linear resource scaling rather than exponential.
© Fraunhofer IKS
Figure 1: Scaling properties of the original surrogation approach versus our proposal (v2.0).
The researchers at Fraunhofer IKShave rigorously tested a lab-developed method, identified its bottlenecks, and created an effective adaptation for practical use (patent application has been filed). This method facilitates the deployment of quantum machine learning solutions into production in a much easier and more cost-effective manner. Our proposal is versatile, with broad applications extending beyond the energy sector, including areas such as healthcare, mobility, and defense.
[1] HSBC, "hsbc.com," 25 September 2025. [Online]. Available: https://www.hsbc.com/news-and-.... [Accessed 26 September 2025].
[2] . J. Schreiber, J. Eisert and J. J. Meyer, "Classical Surrogates for Quantum Learning Models," Phys. Rev. Lett., p. 100803, 2023.
[3] P. A. Hernicht, A. Sakhnenko, C. O'Meara, G. Cortiana and J. M. Lorenz, "Enhancing the Scalability of Classical Surrogates for Real-World Quantum Machine Learning Applications," arXiv:2508.06131, 2025.
