AI-driven models for Cold Spray deposition: transforming additive manufacturing for sustainability Alessia Auriemma Citarella*,† , Luigi Carrino2 , Fabiola De Marco1 , Luigi Di Biasi1 , Alessia Serena Perna2 , Genoveffa Tortora1 and Antonio Viscusi2 1 University of Salerno, CAIS Lab, Department of Computer Science, Italy 2 Department of Chemical, Materials and Production Engineering, University of Naples Federico II, Piazzale V. Tecchio 80, 80125 Napoli, Italy Abstract The integration of Artificial Intelligence (AI) techniques holds promise for advancing the optimization of industrial processes, such as the use of Cold Spray (CS) in the field of Additive Manufacturing (AM). This paper explores the intersection of AI and Cold Spray technology, highlighting its potential to enhance various aspects of AM, including material deposition, surface properties, and process efficiency. Through the utilization of Machine Learning (ML) and Deep Learning (DL) techniques, AI facilitates the analysis of vast datasets encompassing parameters such as powder properties, substrate characteristics, and process conditions, thereby enabling the identification of optimal deposition strategies. Furthermore, AI-driven predictive models offer insights into the complex interactions between process variables, leading to improved understanding and control of the CS process. Keywords Artificial Intelligence, Coldp Spray, Manufacturing 1. Introduction goods, by enabling rapid prototyping, on-demand produc- tion, and the creation of complex structures impossible Industry 4.0, often referred to as the Fourth Industrial with conventional methods. AM offers benefits such as Revolution, represents the integration of digital technolo- reduced material waste, faster time-to-market, and the gies into industrial processes to create smart factories and ability to produce lightweight and optimized components enable more efficient production systems. This transfor- [2]. As it continues to advance, AM holds the potential to mation involves the use of advanced technologies such reshape the future of manufacturing by offering greater as the Internet of Things (IoT), Artificial Intelligence (AI), design freedom, cost efficiency, and sustainability. Sus- robotics, big data analytics, and cloud computing to en- tainability is now imperative in modern manufacturing, hance automation, connectivity, and data exchange in responding to urgent global concerns about environmen- manufacturing. Industry 4.0 aims to improve productiv- tal degradation and decreasing resources. AM can en- ity, flexibility, and customization while reducing costs hance material utilization, reduce environmental foot- and resource consumption [1]. A crucial aspect of Indus- prints throughout product lifecycles, and enable superior try 4.0 is Additive Manufacturing (AM) a transformative engineering functionalities compared to conventional manufacturing process that builds objects layer by layer methods. This holds potential for significant time and from digital designs. Unlike traditional subtractive meth- cost reductions in producing custom [3]. Advancing sus- ods, which remove material from a solid block, additive tainability in AM demands a comprehensive approach manufacturing adds material precisely where needed, al- that extends beyond technical aspects. lowing for intricate geometries and customization. This Among additive technologies, Cold Spray (CS) is gaining technology has revolutionized various industries, from increasing attention owing to its distinctive attribute as aerospace and automotive to healthcare and consumer a cold or non-thermal process, facilitating the treatment of a wide array of materials, including those sensitive to Ital-IA 2024: 4th National Conference on Artificial Intelligence, orga- temperature fluctuations, such as nano-crystalline metals nized by CINI, May 29-30, 2024, Naples, Italy * Corresponding author. or amorphous materials. † These authors contributed equally. Integrating AI models with AM represents a crucial $ aauriemmacitarella@unisa.it (A. Auriemma Citarella); advancement in the evolution of Industry 4.0, offering luigi.carrino@unina.it (L. Carrino); fdemarco@unisa.it substantial benefits in quality control and beyond. By us- (F. De Marco); ldibiasi@unisa.it (L. D. Biasi); ing AI models such as Machine Learning (ML) and Deep alessiaserena.perna@unina.it (A. Perna); tortora@unisa.it Learning (DL) models, manufacturers can optimize pro- (G. Tortora); antonio.viscusi@unina.it (A. Viscusi)  0000-0002-6525-0217 (A. Auriemma Citarella); duction processes, predict potential defects, and ensure 0000-0003-4285-9502 (F. De Marco); 0000-0002-9583-6681 consistent product quality [4]. These models can analyze (L. D. Biasi); 0000-0003-4765-8371 (G. Tortora) vast amounts of data collected from sensors and cam- © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings eras embedded within AM systems, enabling real-time substrate, thereby ensuring the integrity of the coating monitoring and proactive adjustments. Specifically, the [5]. process of CS remains partially manual and uncontrolled. The CS process is governed by several factors. The AI models offer advantageous roads to overcome this size and composition of the powder particles determine challenge by facilitating experimentation and expediting the characteristics of the deposited layer and its adhe- the integration of Low-Pressure CS into industrial work- sion to the substrate. Fine-tuning these parameters can flows. Additionally, AI-powered predictive maintenance optimize material properties and deposition efficiency. can minimize downtime and maximize operational effi- The selection of gas type, flow rate, and pressure controls ciency, further driving sustainability by reducing waste the acceleration and direction of the powder particles and energy consumption. As such, the synergy between during spraying. Adjusting these parameters influences AI and AM can revolutionize manufacturing practices coating density, adhesion strength, and deposition qual- and propel us toward a more sustainable future. ity. The design and geometry of the cold spray nozzle play a vital role in directing the gas-particle mixture onto the substrate [6]. Optimizing nozzle parameters such as 2. Research Fields diameter, shape, and exit velocity ensures precise con- trol over deposition conditions and coating morphology. The primary objective of our research is to design an The distance between the CS nozzle and the substrate, innovative methodology for computer-aided manufactur- known as the standoff distance, affects particle veloc- ing (CAM) systems by leveraging ML and DL approaches ity and impact energy. Adjusting this parameter opti- to develop ad-hoc models. These models attempt to use mizes coating thickness, uniformity, and surface finish cutting-edge technology to discover the best combina- [7]. By carefully adjusting these input parameters, man- tion of factors for CS coating techniques. The advanced ufacturers can tailor the cold spray process to produce parameter sensors used are several such as temperature, high-quality coatings with desired material properties pressure, gas velocity, and optical and auditory sensors. and performance characteristics, all while optimizing for Our research aims at the creation of automated decision- time-consuming and cost efficiency. Fine-tuning param- making systems for monitoring and control, allowing eters such as particle size, gas flow rate, nozzle design, for the deployment of accurate spray strategies. By con- substrate temperature, standoff distance, and powder ducting experiments using a CS machine in real-world feed rate not only ensures the quality of the deposited situations, our goal is to evaluate the effectiveness of layer but also minimizes production time and resource our suggested ML approaches and obtain valuable input usage. This emphasis on efficiency is critical for meeting to improve the accuracy of our forecasting results. The production schedules, reducing manufacturing costs, and findings obtained from this research have the capacity to improving overall competitiveness in the market. improve sustainability and make a valuable contribution However, achieving the ideal combination of parameters to the scientific literature by increasing the effectiveness and selecting their right combinations can be a complex and precision of cold spray coating processes. and time-intensive task. This is where the use of AI mod- These activities involve a multidisciplinary team from els comes into play. By leveraging AI algorithms and the University of Salerno and the University of Naples machine learning techniques, manufacturers can analyze Federico II. The CAIS Lab 1 , laboratory from Computer the amounts of data to identify the most effective parame- Science Department of the University of Salerno, provides ter settings more quickly and accurately than traditional important support in developing algorithms that require methods. This not only streamlines the optimization pro- significant time and resources. Meanwhile, the University cess but also enhances the overall efficiency of the cold of Naples Federico II focuses its efforts on the thorough spray deposition, leading to significant time and cost sav- collection and organization of data. ings while maintaining or even improving the quality of the final product. 3. Cold Spray 3.1. ML models for CS process CS is an emerging technology for micrometer-sized pow- der deposition, increasingly utilized in additive manufac- optimization turing for creating individual components and repairing CS, utilizing kinetic energy and operating at tempera- damaged parts. Among the advantages of CS is the fact tures well below the melting point of metallic particles, that it mitigates thermal degradation and facilitates effi- presents a promising avenue for enhancing the surface cient deposition between the sprayed material and the properties of polymers. While extensively explored and utilized on metal substrates, the application of this tech- nology to polymers remains relatively uncharted terri- 1 https://caislab.di.unisa.it tory, and the underlying physics are not yet fully eluci- dated. This is particularly significant because the charac- The output parameters under consideration are the teristics of the final coating, such as powder deformation penetration depth of the particle and the degree of flat- or penetration depth, crucial for coating adhesion, are tening. In CS, penetration depth refers to the distance influenced by a multitude of factors. These factors in- into the substrate material that the sprayed particles can clude the properties of both the metallic powder and penetrate and adhere to. This depth depends on various the polymeric substrates, alongside the specific spray- factors such as particle velocity, temperature, and ma- ing parameters employed in the process. Consequently, terial properties. Flattening, on the other hand, refers accurately predicting the behavior of metallic particles to the deformation of the sprayed particles upon impact upon impact on various substrates remains a challenge. with the substrate surface. It describes how much the Developing a physical model capable of accurately rep- particles spread out and flatten upon hitting the substrate. resenting the deposition processes of metallic particles Flattening is influenced by parameters like particle ve- onto nonmetallic substrates proved impractical. Validat- locity, temperature, particle size, and substrate material ing such models would require the collection of extensive properties. testing scenarios and outcomes using advanced tools, in- We employed three distinct ML models: Linear Regres- cluding sensors and high-speed cameras. sion (LR), Gaussian Process Regression (GPR), and Neural In this context, Machine Learning (ML) offers the po- Networks (NNs). LR is a straightforward technique uti- tential to reduce the number of necessary experimental lized to establish a linear relationship between variables. trials. However, when we used ML solutions, achieving This relationship elucidates the functional link between precise predictions requires feeding the model with ac- the independent and dependent variables within a given curate and a large amount of data. Therefore, a viable dataset. Consequently, LR models the unknown or de- approach could involve training the model with a com- pendent variable as a linear equation based on the known bination of precise yet limited experimental data and or independent variable. Gaussian Process Regression computational data obtained from Finite Element models (GPR), on the other hand, is a nonparametric Bayesian (FEM), which, while less precise, do not suffer from the approach employed for regression tasks. It excels par- limitations of experimental data. Consequently, in our ticularly with smaller datasets and infers a probability work [8], we used a training dataset composed by 30% distribution model. The Neural Network (NN) used in our of experimental data and 70% of FEM data (mixed data) study is a three-layer function-fitting model trained com- and a second dataset with only FEM data to train some prehensively on the entire dataset. Following training, it ML models. The test dataset for both suggested models can adeptly generalize an input-output relationship. To is fully made up of experimental data. Polyether-ether- assess the efficacy of the ML techniques employed, we ketone (PEEK) and Acrylonitrile butadiene styrene (ABS) computed several performance metrics including Root- substrates were chosen for the deposition process, includ- Mean-Square Error (RMSE), R-Squared, Mean Squared ing both unreinforced and long carbon fiber-reinforced Error (MSE), and Mean Average Error (MAE). Figures 1-2 variants. Spherical powders of copper, aluminum, and report the results for penetration depth and flattening steel were supplied by LPW South Europe for this pur- predictions. The top models for the penetration on the pose. Depositions were carried out using low-pressure test set are NN and LR for the flattening prediction for cold spray equipment (DYCOMET). The samples were po- the mixed data. sitioned on a platform, with the spraying gun mounted on By examining the results derived from training the a robot (HIGH-Z S-400/T-CNC-Technik), which operated model with FEM data, we were able to determine that remotely and sprayed perpendicular to the substrates. the most accurate predictions on the test set for penetra- These materials were chosen for their diverse properties tion and flattening were achieved using the GPR method. and suitability for CS applications. Specifically, the GPR model demonstrated enhanced pen- The input parameters for the strategies employed can etration values while exhibiting a decrease in flattening be categorized into three main groups: performance. Figure 3 depicted the comparison between the performance of the models. • impact velocity, which encompasses other process The conducted experiments indicate that the accuracy parameters such as temperature and pressure; of the models improves with an increasing amount of • powder parameters (Yp), representing the yield available data. Specifically, the models demonstrate bet- strength of the powder material; ter fit to the data when trained on the mixed dataset. By • substrate parameters (Ys), indicating the yield offering an initial insight into the influence of parameters strength of the substrate material. affecting coating deposition, the integration of ML seems The latter also take into account the presence of fibers. to contribute to the optimization of the CS process. When positioned appropriately beneath a matrix layer similar in size to the powders, these fibers reinforce the substrate, modifying its yield strength. pel the field of CS towards greater efficiency, efficacy, and versatility. In this study [9], we had two primary objectives: • firstly, to investigate various DL models aimed at augmenting automation capabilities within the domain of CS; • secondly, we explored the employment of a ge- netic algorithm approach to refine the aforemen- tioned DL models, with the specific aim of en- hancing coating properties. In this scenario, when evaluating potential substrate materials, our study examined a range of options includ- ing ABS, PEEK, and Polyamide PA66 (PA66). Further- more, in selecting powders for the deposition process, we incorporated a variety of metallic options such as copper, aluminum, steel, and titanium. During the initial training phase, we employed two distinct types of datasets to de- Figure 1: Results for penetration depth and flattening predic- velop our models. The first dataset, known as the mixed tions on mixed data dataset, comprised a blend of 30% experimental data and 70% FEM data. Additionally, we utilized a second dataset consisting solely of FEM data, providing a more focused exploration of simulated scenarios. Subsequently, during the evaluation phase, our trained models tested using exclusively experimental data. As input and output pa- rameters, we used the same of our previous work [8]. We employed DL models, specifically focusing on neural network models. These neural networks, inspired by the structure and function of the human brain, are adept at learning complex patterns and relationships in the data. In the development of these networks, we adopted a ge- netic algorithm approach. Genetic algorithms (GAs) draw inspiration from biological evolution, employing princi- ples of natural selection and genetic recombination to find optimal solutions for complex problems. Tradition- ally, GAs are utilized to optimize algorithm parameters; however, in our specific scenario, we employed GAs to de- sign the architecture of the networks. The process of GAs requires several steps: 𝑖𝑛𝑖𝑡𝑖𝑎𝑙𝑖𝑧𝑎𝑡𝑖𝑜𝑛, where potential solutions are randomly selected; 𝑠𝑒𝑙𝑒𝑐𝑡𝑖𝑜𝑛, which iden- tifies optimal parents based on their fitness; 𝑐𝑟𝑜𝑠𝑠𝑜𝑣𝑒𝑟, Figure 2: Results for penetration depth and flattening predic- where genetic material is recombined to generate new tions on FEM data solutions; mutation, introducing random changes to gen- erate genetic diversity; and 𝑒𝑣𝑎𝑙𝑢𝑎𝑡𝑖𝑜𝑛, which assesses the fitness of the solutions. Each of these steps plays a 3.2. Genetic Algorithm-Driven DL models: crucial role in guiding the evolutionary process towards evolutionary approaches to improve identifying optimal solutions for complex problems. In CS deposition this work, we presented two NN as best solutions: Wide Neural (WNN) and Trilayered Neural (TNN) Networks. In our point of view to elevate the deposition efficiency of WNN refers to an artificial neural network architecture CS across a spectrum of materials and to explore the the that typically has fewer hidden layers but a substantial use of AI models, our attention focused on the integration number of nodes in each layer. The TNN is a form of of DL models. Through the integration of DL models, artificial neural network, also known as a single-layer we aim to unlock new insights and capabilities that pro- perceptron, consisting of the input layer, one hidden Figure 3: Comparison of the models Figure 4: Results flattening (on the left) and penetration depth (on the right) predictions layer and the output layer. To assess the effectiveness of prior findings. the DL approaches, we computed several performance metrics, including the RMSE, R-squared, MSE, and MAE. For the flattening, the best results are obtained by the WNN. For the penetration depth, TNN reached the best values, but, except for the R-squared, all the values, on the test set, are high. Overall, the DL models achieved best performance for flattening (see Figure 4). In our previous study, we evaluated ML algorithms using a mixed dataset. NN model excelled in predict- ing penetration on the test set, while LR proved optimal for prediction of flattening. To further comprehend our approach, we compared the outcomes of our earlier in- vestigation with those obtained through the integration of GAs and we reported these results in Table 1 and graphically in Figure 1. Interestingly, the WNN demon- strated exceptional performance in predicting flattening, showcasing a marked improvement over previous results. Figure 5: Comparison between ML and DL models However, it is important to note that for penetration prediction, the TNN constructed with the use of GAs The conducted experiments underscore the potential exhibited a less favorable performance compared to our Table 1 Comparison between ML and DL models Output Models RMSE R-Squared MSE MAE LR 1.83 0.90 3.36 1.45 Flattening WNN 0.73 0.92 0.54 0.51 NN 0.58 0.96 0.34 0.41 Penetration TNN 2.57 0.60 6.64 1.46 of DL techniques in predicting optimal parameter com- neural networks, Journal of Thermal Spray Technol- binations, enhancing the efficiency and effectiveness of ogy (2024) 1–14. the coating process. With the introduction of the GAs [8] A. S. PERNA, L. Carrino, C. A. Auriemma, D. M. Fabi- to improve the design of networks, we can streamline ola, D. B. Luigi, T. Genoveffa, A. Viscusi, et al., A ma- the optimization of model architectures, reducing the chine learning approach for adhesion forecasting of need for manual hyperparameter tuning, which is often cold-sprayed coatings on polymer-based substrates, time-consuming and suboptimal. Materials Research Proceedings 28 (2023) 57–64. [9] A. S. PERNA, L. Carrino, C. A. Auriemma, D. M. Fabiola, D. B. Luigi, T. Genoveffa, A. Viscusi, et al., 4. Acknowledgments Artificial intelligence approaches for enhanced coat- ing performance, Materials Research Proceedings This studies received funding from the European Union - (????). Next-GenerationEU - National Recovery and Resilience Plan (NRRP) – MISSION 4 COMPONENT 2, INVESTI- MENT N. 1.1, CALL PRIN 2022 D.D. 104 02-02-2022 – (OPTIMA: depOsition of cold sPray in the realm of green addiTIve manufacturing through construction of MA- chine learning models) CUP N.D53D23017410001. References [1] H. Lasi, P. Fettke, H.-G. Kemper, T. Feld, M. Hoff- mann, Industry 4.0, Business & information systems engineering 6 (2014) 239–242. [2] A. Bandyopadhyay, S. Bose, Additive manufacturing, CRC press, 2019. [3] H. Hegab, N. Khanna, N. Monib, A. Salem, Design for sustainable additive manufacturing: A review, Sustainable Materials and Technologies 35 (2023) e00576. [4] A. Pratap, A. Pandey, N. Sardana, Machine learning and additive manufacturing: A case study for quality control and monitoring, in: Modern Materials and Manufacturing Techniques, CRC Press, 2024, pp. 211– 234. [5] S. Yin, N. Fan, C. Huang, Y. Xie, C. Zhang, R. Lupoi, W. Li, Towards high-strength cold spray additive manufactured metals: Methods, mechanisms, and properties, Journal of Materials Science & Technol- ogy (2023). [6] R. Dykhuizen, M. Smith, Gas dynamic principles of cold spray, Journal of Thermal spray technology 7 (1998) 205–212. [7] Y. Meng, H. Saito, C. Bernard, Y. Ichikawa, K. Ogawa, Optimal design of a cold spray nozzle for inner wall coating fabrication by combining cfd simulation and