Machining is a key manufacturing technology, representing one of the most significant German economic sectors. To ensure required high-quality assurance and prevent manufacturing errors, process simulations based on digital twins can be applied. However, the current digitization and simulation models face limitations in terms of computational requirements and expert knowledge. As a consequence, important physical efects in industrial practice are either neglected or roughly approximated, resulting in compromised decision-making and economic disadvantages. Since quantum computing (QC) has shown promising benefits in solving numerous algorithmic problems and simulations, the QUASIM research project aims to use QC to improve simulations in manufacturing, reduce modeling eforts and error rates, and develop innovative solutions.
CEUR ceur-ws.org
Metalworking is the largest industrial sector in the European Union [
Early investigations show that quantum mechanical functional principles have decisive
advantages in solving numerous algorithmic problems, resulting in significant accelerations
in numerical procedures, primarily through “Quantum Machine Learning” based approaches
[
Next, we will present descriptive project information, describe the project in more detail (section 2) and provide an overview of the current state of the project (section 3). We conclude the paper by elaborating on limitations, challenges and future work in the project.
The objective of QUASIM is to develop and advance QC-driven approaches that can address critical simulation challenges in manufacturing. QUASIM aims to systematically integrate these QC methodologies into Industry 4.0 frameworks as QaaS solutions, facilitating a seamless transition from problem statements to simplified representations of actual manufacturing systems. Using diferent QC algorithms and methodologies, QUASIM will explore the acceleration and optimization of manufacturing processes through QC-based simulations. The project’s aim includes reducing the economic disadvantages by improving the manufacturing quality and reducing complexities in the simulation. Collaboration with project partners allows the utilization of expert knowledge in exploring hardware capabilities and their limits, ensuring practical quantum adaptation for real-world industrial applications. The project strives to facilitate knowledge transfer for production-oriented simulation based on QC. Results and basic application possibilities of QC are communicated to mechanical engineering companies via the
lation in manufacturing. 01/01/2022 – 12/31/2024 German Research Center for Artificial Intelligence GmbH (DFKI) (Saarbrücken), Jülich Research Center (FZJ) (Jülich), Fraunhofer Institute for Production Technology (IPT) (Aachen), ModuleWorks GmbH (MW) (Aachen), TRUMPF Machine Tools GmbH + Co. KG (Ditzingen) Ford-Werke GmbH (Cologne), MTU Aero Engines (Munich) QUASIM is funded by the German Federal Ministry of Economics and Climate Protection (BMWK) within the funding program ”Quantum Computing - Applications for Industry”, managed by the project management agency German Aerospace Center (DLR) approx. 5.2 million € https://www.quasim-project.de/ industry networks of the partners within workshops, symposia and fairs. Overall, QUASIM aims to transform manufacturing and machining operations and build the foundation for QC’s broader adaptation in the industrial landscape. The project team focuses on two use cases. Within the first use case, whose results are described exemplarily for the project, the target is to improve simulations in the context of laser-cutting. During the cutting process, the heat influence of the laser can lead to unwanted expansion of the material, resulting in an unsuccessful cutting process and machine downtime. We aim to provide QC-based approaches (QC + ML) to improve current simulations and better capture potential thermal expansion in advance of the cutting process. The second use case deals with the development of QC + numerics methods to capture vibrations in the simulation of blade integrated disks (blisks) milling.
Figure 1 represents the target image for the QUASIM project capturing the key components of the system: the Hardware Layer supporting the QC and ML infrastructure, the Platform Layer having platforms to facilitate the AI and QC algorithms, the Algorithmic Layer with classical and quantum-based approaches along with its evaluation, the Data Layer handling simulation input and output data assets, the Service Layer representing the service on the user end where the solutions will be integrated as QaaS, and the Organizational Layer consisting of the stakeholders. Users interact with the system through the Service Layer to request simulations (thermal expansions), which utilize simulation input data (geometry, material properties, etc.) and run QC and classical ML algorithms in the Algorithmic Layer. The results are then handed to the Data Layer and made available to users through the Service Layer.
Table 2 presents the division of the project into diferent work packages (WP), including the
leading institutions of the WP. Each WP is conducted with at least three to four of the project
partners. WP1 and WP9 are complementary WPs, ensuring appropriate project management
and dissemination of results, e.g., through developing demonstrators, and priming of future
users of QC in manufacturing and scientific publishing. In WP2 the needs of the future users are
analyzed and requirements for QC-supported services for industrial production are specified.
WP2 aims to create conceptual models to create abstract representations of the system. It
starts in parallel with the technical modeling (WP3) and the development of the QC+ML and
QC+numerical approaches (WP4 and WP5). Semantic and structural data modeling using
ontologies [
Conceptual models attempt to capture requirements to create a shared understanding among various individuals during the design of a project within the boundaries of the application domain or an organization [14]. For information systems based on programming and machine learning approaches, conceptual models are used as requirements for implementations [15]. In QUASIM, we use conceptual modeling in several contexts. First, we use it in requirements engineering in order to capture the as-is situation in our use cases and to conceptualize the envisioned service system. In addition, we make use of ontologies for representing the data, e.g., in the laser-cutting use case. Therefore, we built on existing ontologies to develop a focused ontology in our project. Furthermore, we incorporate technical modeling for data generation.
In this section, we will explore and describe various aspects of the laser-cutting use case. We will discuss requirements engineering and conceptual modeling, as well as the process of generating synthetic data and data representation. Additionally, we will investigate and evaluate diferent concepts and approaches used for predicting heat flow and thermal expansion. Lastly, we will provide an overview of the current prototype of QUASIM.
The project team followed a thorough requirements engineering process with four steps: (1) Analysis of manufacturing and simulation processes, (2) identification and prioritization of problem spaces, (3) formulation of potential solutions, and (4) derivation of requirements [16]. In the context of (1), we first represented the current service in the as-is situation (see Figure 2) and extended the service design based on the gathered requirements. This conceptual model helps visualize user interactions and corresponding backend actions. It illustrates various user input options like geometry, material properties, and simulation model choices, as well as outputs such as visualization of predicted simulation, and evaluation of model performance. The conceptual model played a crucial role in QUASIM by providing a clear representation of the system, enabling efective communication and informed decision-making, guiding design and development, identifying challenges, and facilitating seamless integration of QC technologies for enhanced simulation capabilities in the manufacturing process.
TRUMPF is currently working on generating real laser-cutting process samples by placing
thermal cameras on their machines to gather heat flow information during the laser-cutting
process. However, to make preliminary studies, we generate synthetic simulations of heat
lfow during a laser-cutting process using Finite Element Method (FEM) [
Here, (, , , ) describes the temperature at point (, , ) at time in a material with thermal conductivity of . The goal is to recover the values of the function (, , , ) which satisfy the conditions of the above PDE. First, we start by modeling the metal sheet geometry in 2-D as a 10cm ×10cm square as seen in Figure 3 (a). Now, we set parameters that are varied throughout the dataset to determine the exact cut geometry i.e. the width of the cutting gap in m ( ), the height of the cut in m (ℎ ), the number of steps into which the cut is discretized ( ). Next, a third dimension is added by turning all 2-D faces into 3-D cuboids with Z-dimension parameterized by depth . Once this 3-D geometry (see Figure 3 (b)) has been obtained, the Matlab meshing algorithm creates a mesh from it (see Figure 3 (c)). Let the mesh have nodes which are connected by the adjacency matrix ∈ ℝ × and have features; ∈ ℝ × . The goal is to train a model which predicts the nodal temperature at a timestep + 1 : ( + 1) ∈ ℝ given input features such as coordinates and temperature at time . Figure 3 (d) shows the simulated temperature on the mesh during laser-cutting, using predictions of a hybrid QML GNN (cf. section 3.4).
For preliminary analysis, we created a large dataset ’ − ’ [18], containing 36 geometries with varying cut widths, heights, and sheet diameters. There are 20 timestamps each per simulation yielding a total of 720 graphs with 800 nodes and 2600 edges on average. In addition, we created a smaller version of this dataset, referred to as ’ − ’ [18], to allow reasonable training times for quantum-based models. − contains 4 geometries with 15 timesteps each, having on average 900 nodes connected by about 3000 edges.
A sub-goal in QUASIM was to develop an approach to represent the data used in the project. We aimed for a semantic description in the sense of an ontology to capture explicit and implicit knowledge, analyze and share the domain knowledge, and enable a clear data representation. We analyzed a multitude of existing ontologies in manufacturing, production, and industry domains, representing a mid-level or domain ontology. Due to its fitting with the domain, generalizability, and rich ontology environment, we chose SAREF 1 as a basic ontology to be extended for QUASIM. We extended SAREF by the classes of “Data Dimension,” “Programmatical Representation,” “Pictures,” “Simulation Backend,” “Classical Backend,” “io” - standing for Input and Output, as well as “Graph,” “Edge,” “Adjacency Matrix,” “Geometries,” “Laser” and “Metal”. An overview of the extended ontology for QUASIM can be found in our GitHub 2.
Inspired by MeshGraphNets [19], we use Graph Neural Networks (GNN) [20] as our base
model architecture [18] given its ability to utilize the local neighborhood of a node to predict
the next temperature making it well-suited for modeling heat difusion spatially for diverse
geometries. We use Multi-Layer Perceptrons (MLPs) [21, 22] to embed node features into a latent
space and decode final output predictions. For the message passing block, we evaluate three
diferent aggregation functions: a) GraphConvolutionalNetwork (GCN) [ 23], b) GraphConv
[24], and c) Graph Attention Networks (GAT) [25]. To explore quantum advantages [
As first results, we show the comparison of the diferent encoding methods: Quantum-based encoding ( ) with () parameters and MLP-based encoding − with ( 2)parameters and −2 with (2 ) parameters for both datasets is shown in Table 3. We observe that the models with quantum embeddings ( ) exhibit the best performance with an MSE of 1https://saref.etsi.org/core/v3.1.1/ 2https://github.com/InformationServiceSystems/quasim-project
ML− ML−2 QML .
GCN [23] GraphConv [24]
GAT [25] 0.21 on LAGER-SMALL. By using a QML-based approach instead of classical MLP, we achieve significantly lower errors with exponentially fewer parameters than competing models. This empirical evidence demonstrates the power of QML models, benefiting real-world applications.
We present a UI demonstrator that showcases heat flow simulations generated by our models during laser-cutting operations. The prototype allows users to select input parameters such as material properties and laser flux, and to visualize and compare predicted heat flow simulations among diferent models. A demo video of the QUASIM prototype can be found in the GitHub. 3
We presented QUASIM, a research project leveraging QC to address critical simulation challenges in manufacturing processes, ofering QaaS solutions for integration into Industry 4.0 frameworks. By combining QC with ML, QUASIM seeks to significantly improve simulation accuracy and reduce computation times, leading to economic benefits, and faster manufacturing processes. The current work on laser-cutting simulations using hybrid QML models shows promising results. With QUASIM, manufacturing companies are introduced to how QC can be practically used to identify and implement competitive advantages, contributing to the future success of mechanical engineering. As the next steps, we plan to investigate fully quantum GNNs for faster and more eficient thermal simulations, as well as implementing and evaluating the TIGNN approach. Despite the challenges of QC’s current state (hardware limitations, scalability, noisy estimations) [34, 35, 36], we are confident that QUASIM through its excellent consortium support and meticulous investigation will provide innovative solutions enhancing manufacturing quality. To overcome the challenges, we investigate how the developed solutions can be used with today’s noisy hardware as well as with perfect simulators, representing the next quantum hardware generation.
This work is part of the research project QUASIM (grant number: 01MQ22001A), funded by the German Federal Ministry for Economic Afairs and Climate Action (BMWK). 3https://github.com/InformationServiceSystems/quasim-project [14] W. Maass, V. C. Storey, T. Kowatsch, Efects of external conceptual models and verbal explanations on shared understanding in small groups, 2011. URL: https://doi.org/10.1007/ 978-3-642-24606-7_8. doi:10.1007/978-3-642-24606-7\{_8. [15] F. Bozyiğit, O. Aktaş, D. Kilinc, Linking software requirements and conceptual models: A systematic literature review, Engineering Science and Technology, an International Journal 24 (2021) 71–82. URL: https://doi.org/10.1016/j.jestch.2020.11.006. doi:10.1016/j. jestch.2020.11.006. [16] H. Stein, S. Schröder, P. Kienast, M. Kulig, Towards requirements engineering for quantum computing applications in manufacturing, in: Hawaii International Conference on System Sciences 2024. Hawaii International Conference on System Sciences (HICSS-2024), HICSS, 2024. [17] T. J. R. Hughes, The finite element method: linear static and dynamic finite element analysis, 1987. URL: http://zetusyz.files.wordpress.com/2014/06/ the-finite-element-method-linear-static-and-dynamic-finite-element-analysis.pdf. [18] S. Xu, F. Wilhelm-Mauch, W. Maass, Quantum feature embeddings for graph neural networks, in: Hawaii International Conference on System Sciences 2024. Hawaii International Conference on System Sciences (HICSS-2024), HICSS, 2024. [19] T. Pfaf, M. Fortunato, A. Sanchez-Gonzalez, P. W. Battaglia, Learning Mesh-Based Simulation with Graph Networks, International Conference on Learning Representations (2021).
URL: https://dblp.uni-trier.de/db/conf/iclr/iclr2021.html#PfaffFSB21. [20] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, G. Monfardini, The Graph Neural Network model, IEEE Transactions on Neural Networks 20 (2009) 61–80. URL: https: //doi.org/10.1109/tnn.2008.2005605. doi:10.1109/tnn.2008.2005605. [21] A. P. Sánchez, J. Godwin, T. Pfaf, Rex, J. Leskovec, P. W. Battaglia, Learning to Simulate Complex Physics with Graph Networks, International Conference on Machine Learning 1 (2020) 8459–8468. URL: http://proceedings.mlr.press/v119/sanchez-gonzalez20a/ sanchez-gonzalez20a.pdf. [22] Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, P. S. Yu, A comprehensive survey on graph neural networks, IEEE transactions on neural networks and learning systems 32 (2021) 4–24.
URL: https://doi.org/10.1109/tnnls.2020.2978386. doi:10.1109/tnnls.2020.2978386. [23] T. N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, in: 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings, OpenReview.net, 2017. URL: https: //openreview.net/forum?id=SJU4ayYgl. [24] C. J. Morris, M. Ritzert, M. Fey, W. L. Hamilton, J. E. Lenssen, G. Rattan, M. Grohe, Weisfeiler and Leman Go Neural: Higher-Order Graph Neural Networks, Proceedings of the AAAI Conference on Artificial Intelligence 33 (2019) 4602–4609. URL: https://doi.org/10.1609/ aaai.v33i01.33014602. doi:10.1609/aaai.v33i01.33014602. [25] P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Liò, Y. Bengio, Graph attention networks, in: 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings, OpenReview.net, 2018. URL: https://openreview.net/forum?id=rJXMpikCZ. [26] R. LaRose, B. Coyle, Robust data encodings for quantum classifiers, Physical review 102 (2020). URL: https://doi.org/10.1103/physreva.102.032420. doi:10.1103/physreva.102. 032420. [27] M. Schuld, R. Sweke, J. J. Meyer, Efect of data encoding on the expressive power of variational quantum-machine-learning models, Physical review 103 (2021). URL: https: //doi.org/10.1103/physreva.103.032430. doi:10.1103/physreva.103.032430. [28] P. Selig, N. Murphy, A. S. R, D. Redmond, S. Caton, A Case for Noisy Shallow Gatebased Circuits in Quantum Machine Learning, International Conference on Re- booting Computing (ICRC) (2021). URL: https://doi.org/10.1109/icrc53822.2021.00015. doi:10.1109/ icrc53822.2021.00015. [29] S. Lloyd, M. Schuld, A. Ijaz, J. Izaac, N. Killoran, Quantum embeddings for machine learning, arXiv (2020). URL: http://export.arxiv.org/pdf/2001.03622. [30] S. Y.-C. Chen, T.-C. Wei, C. Zhang, H. Yu, S. Yoo, Hybrid Quantum-Classical Graph Convolutional Network, arXiv (Cornell University) (2021). URL: https://arxiv.org/pdf/2101. 06189.pdf. [31] M. A. Nielsen, I. L. Chuang, Quantum Computation and Quantum Information, 2012. URL: https://doi.org/10.1017/cbo9780511976667. doi:10.1017/cbo9780511976667. [32] Q. Hernández, A. Badías, F. Chinesta, E. Cueto, Thermodynamics-informed graph neural networks, IEEE transactions on artificial intelligence (2022) 1. URL: https://doi.org/10. 1109/tai.2022.3179681. doi:10.1109/tai.2022.3179681. [33] Z. M. Ruhi, H. Stein, W. Maass, A proposal for physics-informed quantum graph neural networks for simulating laser cutting processes., 2023. [34] M. Amir, C. Bauckhage, A. Chircu, C. Czarnecki, C. Knopf, N. Piatkowski, E. Sultanow, What can we expect from quantum (digital) twins?, 2022. URL: https://publica.fraunhofer. de/handle/publica/430446. [35] F. Bova, A. Goldfarb, R. G. Melko, Quantum Economic Advantage, NBER Working Papers 29724, National Bureau of Economic Research, Inc, 2022. URL: https://EconPapers.repec. org/RePEc:nbr:nberwo:29724. [36] W. Chipidza, Y. Li, A. Mashatan, O. Turetken, L. Olfman, Quantum computing and is harnessing the opportunities of emerging technologies., 2023. doi:https://doi.org/10. 17705/1CAIS.05219.