Submitted Successfully!
To reward your contribution, here is a gift for you: A free trial for our video production service.
Thank you for your contribution! You can also upload a video entry or images related to this topic.
Version Summary Created by Modification Content Size Created at Operation
1 -- 5522 2023-06-08 20:21:52 |
2 format Meta information modification 5522 2023-06-09 02:59:40 | |
3 format -406 word(s) 5113 2023-06-15 05:32:10 |

Video Upload Options

Do you have a full video?


Are you sure to Delete?
If you have any further questions, please contact Encyclopedia Editorial Office.
Jenis, J.; Ondriga, J.; Hrcek, S.; Brumercik, F.; Cuchor, M.; Sadovsky, E. Machine Learning in Mechanical Design and Optimization. Encyclopedia. Available online: (accessed on 17 June 2024).
Jenis J, Ondriga J, Hrcek S, Brumercik F, Cuchor M, Sadovsky E. Machine Learning in Mechanical Design and Optimization. Encyclopedia. Available at: Accessed June 17, 2024.
Jenis, Jozef, Jozef Ondriga, Slavomir Hrcek, Frantisek Brumercik, Matus Cuchor, Erik Sadovsky. "Machine Learning in Mechanical Design and Optimization" Encyclopedia, (accessed June 17, 2024).
Jenis, J., Ondriga, J., Hrcek, S., Brumercik, F., Cuchor, M., & Sadovsky, E. (2023, June 08). Machine Learning in Mechanical Design and Optimization. In Encyclopedia.
Jenis, Jozef, et al. "Machine Learning in Mechanical Design and Optimization." Encyclopedia. Web. 08 June, 2023.
Machine Learning in Mechanical Design and Optimization

Machine learning, by definition, refers to the development of knowledge by computers using input data. Mechanical engineering is a strong area for this technology. Older machine builders have gained significant knowledge through several case studies and efficient workflows. This creates the opportunity to approach new projects, to be able to either use their technical knowledge with familiar elements that have already been discovered or use their experience to learn and adapt to new challenges. Newcomers to the field learn in this way and gather their effectiveness and knowledge; when it comes to learning, engineering is one of the areas where much of the learning takes place at work. How well it pays depends on the ability to collect and process information.

artificial intelligence machine learning deep learning

1. Product Design

Machine learning can improve the way machines and equipment are designed in many ways, e.g., by combining different parallel analyses of physics, solid mechanics, fluid mechanics, and so on. Using neural networks, it is possible to force a computer to learn to distinguish between parts and different assemblies and tools, which can be used in the future for designing a wide range of components and for data entry purposes in the mass production of various products. The following subchapters provide more detailed research projects from around the world, which deal with integration, respectively, using artificial intelligence and machine learning in the process of mechanical construction and design.

1.1. Platform for Integrated Aircraft Design

Efforts to shorten aircraft design times have stimulated the development of collaborative frameworks that facilitate the design of aircraft and their systems in a more integrated manner. The objective of the project was to create an integrated modeling and simulation framework that combines a filtering process to reduce the number of feasible architectures, a modeling platform that simulates an aircraft’s power system, and a machine learning-based clustering and optimization module. This framework enables the designer to prioritize several designs and provides traceability for the best choices. In addition, it allows the integration of models at several degrees of realism, depending on the size of the design area and the desired precision. It exhibits the use of different electrical control technologies to electrify the main flight control system (PFCS) and the landing gear braking system. Key performance indicators are used to assess the performance of various architectures (fuel consumption, weight, performance). The optimization procedure employs a data-driven localization stage to recognize files with comparable structures. The framework displays the capacity to optimize across numerous system architectures in a manner that is efficient and scalable for bigger design spaces and problems of greater sizes [1].

1.2. Investigate Models for Efficient Decoding of 3D Point Clouds

This focuses on surrogate modeling strategies to learn the approximation of computationally demanding 3D model function assessments. In the past, 3D point clouds were considered too large as data formats for surrogate modeling. However, employing neural network advancements to automatically encode 3D objects, these point clouds may now be translated to one-dimensional latent spaces [2]. This gives rise to a fundamental research question: what alternative modeling approach is best for discovering the links between the 3D geometric attributes of objects represented in a coded latent vector and the physical events collected in the assessment software? Unintuitively, it has been shown that decoding latent representations of 3D objects into performance predictions is far more successful than using a neural network decoder. Surrogate models without neural networks obtain equivalent accuracy to neural network models in test instances, including 3D model airplane datasets and watercraft datasets. A test case confirmed the equivalent accuracy of the modeling methodologies, but it was also discovered that the distribution of data performance values, particularly the existence of numerous outliers, had a considerable negative influence on accuracy. These findings challenge the commonly held belief that neural networks provide an effective “universal” solution for learning black functions and imply that even in systems that use several neural networks, possibly more efficient solutions for each network should be examined. Depending on the needed application accuracy, this surrogate modeling technique might be used to simulate costly simulation software, or if the error tolerance is low, it can be used as a first step to restrict the number of conceptual designs for further examination [3][4].

1.3. Machine Learning of Object Shapes through 3D Generative-Adversarial Modeling

A novel framework, the 3D generative inverse network (3D-GAN), which creates 3D objects from probability space utilizing recent breakthroughs in volume convolutional networks and generative reverse networks, was suggested. The proposed model has three advantages: first, the use of an inverse criterion instead of traditional heuristic criteria enables the generator to implicitly capture the structure of the object and enables the synthesis of high-quality 3D objects; second, the generator creates mappings from low-dimensional probability space to 3D object space, making it possible to sample objects without a reference image or CAD model and examine the diversity of 3D objects; and third, the proposed model is applicable to a wide range of applications. Experiments demonstrate that the suggested technique creates high-quality 3D models, and that the learned attributes (unsupervised machine learning) achieve excellent performance in identifying 3D objects, equivalent to the performance of supervised learning methods [5].

1.4. Design of a Spatial Recurrent Neural Network for Design Form Optimization

A novel strategy for optimizing the structure and behavior of complex systems was devised. The method employs spatial grammar contained in character-recurrent neural networks (RNNs) to define the system, including the number of actuators and degrees of freedom, reinforcement learning to optimize actuator behavior, and physical simulation systems to assess performance and provide data for RNN (re)training. Grammar and RNN allow for a more complicated, combinatorically limitless design space as compared to the parametric optimization of a design with a fixed number of inputs. In the suggested technique, the RNN is first taught the spatial language defining the assembly arrangement, component geometry, material attributes, arbitrary numbers, and actuator degrees of freedom. The produced designs are further assessed in a physics-based environment, with the internal optimization loop using improved learning to identify the optimal actuator control strategy [6]. An RNN equipped with high-performance grammar can generate a design that is optimized in terms of both shape and behavior. Two assessment case studies based on the design of a modular sailing vessel are provided. The first case study optimizes the design without control surfaces, enabling RNN to comprehend the rationale behind high-performance solutions. The second case study expands the first to include controllable actuators, necessitating improvement of the inner loop’s behavior [7].

1.5. Use of Construction Space for the Design of Additive Production

Through logical and integrated product designs, materials, and production methods, this project’s additive manufacturing design maximizes product performance. It is difficult to find design solutions in such a complex design area. It is notable that no current design assistance solution is both quick and tailored to the design process. A holistic strategy for designing and optimizing the design across the various phases of the design process is provided. In particular, a two-step design process based on different models for the execution and detailed design stages is offered. The Bayesian network classifier is utilized as the reasoning framework for investigating the design space during the design phase, whilst the Gaussian process regression model is employed as the evaluation function for the optimization approach to utilize the design (construction) space during the detailed design. On the basis of a single dataset generated by the Latin hypercube sampling technique, and improved using the Monte Carlo Markov chain sampling method, these models were developed [8]. This cost-effective data-based method is exemplified in the design of a customizable ankle brace that has mechanical characteristics that can be tuned utilizing a high-expansion design concept with bespoke stiffnesses [9].

1.6. Generative Design and Verification of 3D Conceptual Wheel Models

This investigation focuses on an engineering design that combines artificial intelligence (AI) with computer-aided design (CAD) and computer-aided engineering (CAE). A CAD/CAE framework based on deep learning was developed for the conceptual design phase, which automatically creates 3D CAD drawings and analyzes their technical performances. The following steps comprise the proposed framework: 2D generative design, dimensionality reduction, latent space experiment design, automated production of 3D CAD models, automatic modeling of CAE findings, transfer learning, visualization, and analysis. On the basis of a case study involving the design of a road wheel, the suggested framework demonstrates that AI may be realistically included into a final product design project [10]. Using this framework, engineers and industrial designers may review a large number of 3D CAD models together with the findings of technical performance (CAE analysis) predicted by AI to determine the ideal conceptual design variations for the future phase of detailed designs [11].

1.7. Machine Learning-Assisted Propeller Design

In Maritime and subsonic aircraft, propellers are some of the most common propulsion mechanisms utilized to create thrust from the rotating motion of the engine. Due to their simplicity, durability, and great efficiency, propellers have been the most popular design options used in the last century [12]. Nonetheless, it remains difficult to identify the best application-specific shape. This addresses the use of contemporary and quickly advancing machine learning (ML) methods in the production of new designs. It leverages a massive collection of previously existing parametric design patterns and includes technical knowledge, with extremely realistic simulation models; thus, the design process was approached as a supervised learning problem.
The objective was to create and test machine learning models for parametric propeller designs using application-specific restrictions. The outcome was the discovery of an unorthodox propeller geometry with extremely high efficiency. The exceptionally high accuracy of validation data is exemplary, and these results suggest that further progress in this direction can help examine new, unrecognized, effective designs and save designers a great deal of time; further progress can also help to alleviate the challenges of validating these designs beyond accuracy [13].

1.8. Evaluation of Steel Structures Using Machine Learning

A novel design method based on an iterative machine learning algorithm was presented to expedite the topological investigation of shell structures with planar surface compression, while taking structural performances and design restrictions into consideration. Neural networking enables the training of a surrogate model to expedite the assessment of the structural performance of a variety of potential structural forms without the need for a substantially slower process of locating geometric shapes [14]. As the principal tool of structural designs, methods of identifying geometric shapes of 3D graphic statics are utilized to produce a single-layer shell by compressing flat surfaces. Under identical boundary conditions, the partitioning of the force diagram and its polyhedral cells using different rules results in topologically distinct structures that possess varying carrying capabilities. The solution space for all potential forms with pure compression under a given boundary condition is enormous, making it very hard to iterate over all forms in order to locate optimal solutions. After training with iterative active sampling, the surrogate model can analyze input data, including distribution rules, and forecast the structural performance and design restrictions of planar surfaces in milliseconds. Consequently, it is feasible to analyze the nonlinear interactions between all distribution rules and the chosen structural performance metrics, and then display the full solution space. As a consequence, a number of design solutions with specific assessment criteria have been discovered, demonstrating the effectiveness of this form-finding technique. Moreover, given the overall training time of the neural network model, the suggested framework is still quicker than a conventional optimization technique, such as a genetic algorithm that can only discover optimum values. This procedure yields interactive sampling approaches in which machine learning models aid the designer in picking and controlling various design strategies by giving real-time feedback on the impacts of chosen parameters on design results [15].

1.9. Visual Search Application for Machine Components

This examined a comprehensive, annotated benchmark of mechanical components for classification and retrieval operations. It is a comprehensive data collection of 3D mechanical component objects. This dataset enables the data-driven study of mechanical component symptoms. Examining the form descriptor of mechanical components is vital for computer vision and industrial applications. However, developing annotated datasets on mechanical components has not received much attention. Obtaining 3D models is difficult because annotating mechanical components requires technical expertise. The primary contributions are the establishment of a large reference collection of annotated mechanical components, the definition of a hierarchical taxonomy for mechanical components, and a comparison of the performances of deep learning classifiers for shape analysis applied to mechanical components. Seven cutting-edge techniques for categorizing deep learning into three categories, namely cloud points, volume representation in voxel grids, and representation-based representation, were examined using an annotated dataset. Autodesk, the developer of Design Graph software, is conducting similar research [16][17].

2. Optimization of Strength Designs

2.1. Machine Learning as a Substitute for the Finite Element Method

Current maintenance intervals for mechanical systems in transportation are based on system life, which results in expensive maintenance planning and often puts passengers at risk. In the future, the actual usage of the vehicle will be utilized to anticipate the stresses in its design, therefore enabling the development of a particular maintenance plan. Machine learning (ML) methods may be used to transfer a subset of real-time design measurements to a finite element analysis-based, high-fidelity model of the same system (FEM). Consequently, the ML technique based on finite elements directly calculates the stress distribution across the system during operation, hence enhancing the capacity to create safe and effective maintenance procedures [18]. An alternative finite element method based on machine learning methods was presented for estimating the time-varying response of a one-dimensional beam. Several regression models, such as decision trees and artificial neural networks, were created; their effectiveness in estimating the stress distribution in the beam structure was evaluated. Substitute finite element models based on ML algorithms may predict the beam response correctly; however, artificial neural networks have produced more precise results [19].

2.2. Application of Machine Learning for the Calculation of Pressure Equipment

The use of finite element analysis (FEA) is widely prevalent in the design of pressure equipment. While the finite element approach offers stress distribution in the geometry of pressure equipment, the analyst must manually specify specific parameters and boundary conditions in order to conduct the analysis. The purpose was to enhance or replace manual analytical processes using machine learning. Two distinct models have been created to replace two similar manual pressure equipment analysis processes. The models were trained using 605 distinct datasets derived from the stress study from a common region of discontinuities in pressure equipment. In order to detect discontinuities and anticipate linearized stresses, machine learning models were created, thereby accelerating the analytical process. The findings indicate that the trained models are sufficiently accurate at greatly augmenting the analytical procedure [20].

2.3. Optimization of Mechanical Micro-Mixer Design Using Machine Learning

Due to the expenses involved with producing and testing multiple prototypes, it is often challenging to establish an ideal multipurpose design in practical engineering applications. Therefore, it is advised to use calculation tools. machine learning approaches (solved as design optimization with machine learning help) were utilized to build a mechanical mixing device consisting of a microchannel-enclosed rectangular column. The random forest classifier was trained to predict which geometries might result in vortex scattering. Next, a multi-objective optimization issue was explored, which consisted of decreasing the needed pumping power and optimizing the mixing efficiency within the restrictions of a given design. If additional trained data were necessary for alternative configurations, the random forest classifier could have been used to predict whether or not it was worthwhile to simulate a new configuration, thereby preventing the simulation of unnecessary computationally intensive cases and accelerating the data-driven process. The performance of the optimum designs derived from the application of the genetic algorithm on surrogate samples was simulated and shown. Even under very unfavorable mixing circumstances and a somewhat low Reynolds number, optimal geometric arrangements give optimal mixing efficiencies at very low short-channel pumping costs, exceeding conventional technology. By using form parameterization methodologies, the MLADO framework that guides may be expanded and automated for other comparable engineering design processes of any size [21].

2.4. Use of Machine Learning for the Prediction and Optimization of Mechanical Systems

This examined the precision of predictions made by multiparametric models obtained from numerical data. There were three distinct mechanical test scenarios utilized to produce the data. Models were constructed from the data to predict changes in attributes in response to arbitrary changes in input parameters. Various modeling methodologies were assessed based on the accuracy of their predictions. By providing machine learning toolboxes, polynomial matrix equations were contrasted with regression models and neural network models. Model similarities and differences were discussed. A matrix-based exponential model was developed to improve the precision of certain applications. The effects and causes of the model prediction accuracies were evaluated. This defines the basic conditions for developing useful models. This resulted in a comparison of modeling methodologies in terms of their physical plausibility and efficacy. There is a correlation between data production and the training process and efficiency. For one of the example situations, a prediction model was used to show its application and capabilities. Using the model equation, the value of the sanction function in an optimization job with various inputs and outputs was calculated. By updating the material parameters, the outcome of the optimization involved the adjustment of the four natural frequencies to the observed values. In every other instance, sensitivity analyses were conducted, including the validation of numerical findings [22].

2.5. Topology Optimization

Recent advancements in design optimization have the potential to significantly enhance the performances of mechanical components and systems. In combination with additive manufacturing, topology optimization is one of the numerical approaches used to algorithmically build optimized designs that influence the mechanical design of existing hardware. Sadly, many of these algorithms may require substantial human configuration and control, particularly the debugging parameters that govern the algorithm’s performance and convergence. This proposes a machine learning-based framework to recommend tuning parameters to users, therefore avoiding the expensive trial-and-error process of human tuning. The program collects the debugging parameters from the repository of past comparable issues, which are assessed using different parameters, depending on the problem information, and refines them using the Bayesian optimization technique for the present problem. The technique is illustrated on a basic topological optimization problem with the objective of achieving a decent topological optimization solution and then determining the appropriate “compromise” between the solution quality and necessary computing time. The objective is to decrease the number of “unnecessary” debugging starts needed for manual debugging alone. With more improvement, this system may eventually be applicable to enterprise-level analysis and optimization problems. Topology optimization is one example, but the framework may also be used for other optimization issues, such as form and dimensioning in highly accurate physics-based analytical models [23].

2.6. Design of Hollow Section Columns Using Machine Learning Methods

Many current techniques for designing the load-bearing capacity of round, hollow, stainless-steel columns have been established for a particular class, taking the global deflection failure approach into consideration. However, various grades of stainless steel have significantly varying material characteristics, and hollow section columns are susceptible to local deflection, global deflection, and both global and local interactive deflections. A framework based on machine learning was employed to build a standard design process applicable to various grades of stainless steel and failure mechanisms. First, 39 experiments on cold-formed hollow section columns were conducted. Material characteristics, flaws, load and deformation curves, and failure mechanisms were described in detail [24]. Then, test data for these stainless-steel columns were gathered, and a database of 280 columns was generated. Then, two machine learning methods, random forest and extreme gradient boosting, were used to forecast column load capacities based on four different kinds of input characteristics. Using all design inputs, the random forest algorithm had maximum prediction accuracy. The inclusion of the yield strength to Young’s modulus ratio as an input parameter considerably enhanced the accuracy of the random forest technique when based on complicated factors (i.e., the dimensionless cross-sectional area and bar slenderness) [25].

2.7. Prediction of Load-Bearing Capacity of Double-Cut Screw Joints

This involves the ground-breaking use of machine learning to forecast the load-bearing capabilities of double-cut structural steel bolted joints. A database containing 443 experimental datasets with input characteristics, such as standardized end distances, edge distances, bolt spacing, transverse to load directions, sheet steel strength ratios, load cell rows, joint configuration, and standard load capacity, was compiled for the first time. For this application, eleven machine learning approaches were studied. The investigation of the significance of the factors revealed that the normalized distances between the ends and edges had the largest influence on the final strength. Using different statistical indicators, the performances of the models were assessed and compared to current formulations and requirements of the proposed legislation. The model with the greatest coefficient of determination, the lowest mean absolute error, and the lowest mean square error was the random forest model. In contrast to previous models, which are particular to individual steel grades and give separate equations for various failure modes, machine learning models have achieved accurate, integrated, and generic approaches to prediction. Intriguingly, these models demonstrate that the ratio of steel tensile strength to compressive strength and the number of screw rows, which are presently disregarded in the design recommendations, considerably influence bearing strength (by about 10%). A user-friendly interface comprising all of the planned machine learning algorithms was created, which helped the design of double-cut screw joints and acted as a teaching and research resource [26].

2.8. Design and Optimization of Pneumatic Gear Drive

Electro-pneumatic drives play a vital role in a variety of sectors, including in the construction of large vehicles. Here investigates the topic of operating an automated manual gearbox whose actuator is a double-acting cylinder with a floating piston and a reserved internal position. When constructing the control of electro-pneumatic cylinders, it is important to execute control with a fixed value on a nonlinear system when the airflow is provided by disproportionate valves, as is the case here. Due to the fact that both the system model and the qualitative management objectives may be expressed as partly observable Markov decision-making processes, machine learning frameworks are obvious options for addressing such management issues. This analyzes six distinct options for this purpose. In a simulated environment, the performances and strategic choices of these six approaches were compared. To illustrate the applicability of the idea, validation tests were conducted on an actual transmission system and implemented on the vehicle’s control unit. In this instance, the policy gradient agent was chosen based on implementation limitations and computing capability. The findings indicate that the given approaches are appropriate for the control of floating piston cylinders and may be applied to other mechanical fluid drives and nonlinear control problems with fixed values [27].

3. Getting Human Preferences and Design Strategies

Recent advances in artificial intelligence offer the perfect opportunity to integrate this technology (including machine learning) into the human design and construction teams. To date, machine learning approaches have not paid enough attention to the consideration of human aspects, which play an important role in the design process. These subchapters present research and studies in the field of obtaining human preferences and strategies, which can be used in decision-making in the design and development of various products.

3.1. Data-Driven Methodology for Creating Customer Selection Sets Using Online Data and Customer Reviews

Recent innovations in technological design include a selection model that takes client preferences into account. When constructing a choice model that enables one to gain an accurate estimate of the parameters associated with product features, a solid collection of options is crucial. However, the selected file information is often unavailable [28]. This entry offers a mechanism for constructing customer samples using web data and customer evaluations in the absence of real samples or sociodemographic consumer data. The approach consists of three primary components, i.e., categorizing items according to their qualities, grouping customers according to their evaluations, and building sample files based on a selection probability scenario based on product and customer groups. The normalized scenario describes the suggested scenario, which multiplies the proportions of product clusters and consumers to generate a probabilistic distribution of choice. A linear combination of product qualities alone and a feature that incorporates the interplay of product attributes and customer reviews are provided as utility functions. The approach was applied to a collection of laptop data. Predicting the test file data, the normalized scenario obtains much better outcomes than the baseline (random) scenario. In addition, the addition of customer evaluations to the utility function considerably improves the model’s predictive ability. Using product attribute data and customer evaluations to produce samples yields sample models with more predictive ability than randomly generated samples, according to research [29].

3.2. Obtaining Customer Insights on Product Sustainability from Online Reviews

For a product to be successful on the market, its designers must produce items that are not only real but also sustainable in terms of consumer perceptions-; both the reality and perceptions of competition may vary significantly. This  details the design technique used for determining the perception of sustainable features via the gathering of online reviews, their human annotations via crowdsourcing, and the processing of review fragments annotated by a natural language machine learning algorithm. As indicated by the observed discrepancy between the perception of sustainability and the actuality of sustainable design, the research demonstrates that a number of critical environmental sustainability problems, such as energy and water use, did not have a substantial influence on consumer sentiment. Based on these outcomes, internet reviews may help designers acquire competitive traits for future design studies and produce products that are compatible with sustainable consumer values [30].

3.3. A Data-Driven Approach to Identify the Context of Product Use from Online Customer Reviews

This entry presents a data-driven technique for identifying product use scenarios automatically from online customer evaluations. The environment of product usage influences product design, user behavior, and consumer happiness. Prior research identified use settings through a survey-based technique or by subjective means. In contrast, the suggested technique employs natural language machine learning and processing methods to discover and aggregate use contexts from a large number of customer evaluations. In addition, aspect analysis is utilized to extract sentiment within a specific phrase context. The outcome demonstrates that the approach is capable of capturing important product usage contexts and clustering bigrams associated with comparable use contexts. Aspect sentiment analysis enables the observation of a product’s position relative to its rivals in a certain context of usage. This insight may highlight to the product creator a need to enhance the product. It may also signal a possible marketing opportunity in an application scenario where the majority of existing items are adversely evaluated by consumers. Due to a minor linear connection for the majority of use contexts in the case study, it suggests that the overall evaluation may not be a significant predictor of the customer’s mood in regard to a specific use scenario [31].

3.4. Mining and Representing the Conceptual Space of Existing Ideas for the Purpose of Targeted Idea Generation

Frequently, design innovation initiatives produce a great number of design ideas from designers, consumers, and, increasingly, regular internet users. Such data are often utilized for selection and execution, but they may also serve as sources of inspiration for the generation of further ideas. Specifically, the basic concepts underlying the original ideas may be recombined to form new concepts. Obtaining concepts from raw lists of ideas and data sources in order to inspire or develop new ideas is not a simple operation. The fact that data on concepts is often presented in unstructured natural languages is a key issue. A methodology is developed that employs natural language processing to extract keywords as elementary concepts embedded in massive descriptions of ideas and represents the space of elementary concepts in a core-periphery structure to facilitate the recombination of elementary concepts into new ideas. The methodology is applied to the extraction and representation of the conceptual space that underlies the massive ideas gained from crowdsourcing and is used to generate new ideas for future transport system designs in a real project funded by the public sector through the use of people and automated computer programs. The investigation into the processes and outcomes of human and machine recombination provides information on the future directions of artificial intelligence research in the domain of design concepts [32].

3.5. Transferring Design Strategies from Human to Computer

Solving every design issue requires planning and developing techniques that identify and prioritize subprocesses. It is a talent that designers acquire through time and then use to tackle comparable situations. This transfer of design techniques has not, however, been successfully modeled or implemented inside computational agents. This proposes a probabilistic model-based method for representing design methods. The model offers a method for generating new designs based on specific design strategies as the configuration design challenge is gradually addressed. This experiment also demonstrates that this probabilistic representation may be used to transmit human designer techniques to computer design agents in a generic and effective manner. This transfer-based method enables the identification of high-performance designer behavior and its application to the management of computational design agents. Lastly, the agents exemplify the core behavior of transfer learning since the transfer of design techniques across challenges has resulted in enhanced agent performance. This employs the CISAT (cognitively inspired simulated annealing teams) framework, an agent-based model that replicates human problem-solving in configuration design challenges [33].

3.6. Mimicking Human Designers through Deep Learning

People, as designers, have rather diverse problem-solving techniques. Computer agents, on the other hand, have access to substantial computational resources for solving specific design issues. Consequently, if agents can learn from human behavior, a synergistic team may be established to handle challenges involving humans and artificial intelligence. Here proposes a method for extracting human design strategies and implicit norms from historical human data and using them to create designs. This involves the creation and implementation of a two-tiered framework that mimics human design techniques via observational learning. This system uses deep learning components to build designs without explicit goals and performance parameter information. The framework is meant to engage with the issue using a visual interface, similar to how individuals solve the challenge. It is taught to imitate a group of human designers by monitoring the sequences of their design states, without introducing a problem-specific modeling interest or extra issue knowledge. In addition, a final agent is constructed that employs this deep learning architecture as its foundation, together with image processing techniques to map pixel motions to the design as a method for design development. Finally, the designs created by the calculation teams of these agents are compared with real human data for the teams tasked with addressing roof truss designs. The findings demonstrate that these agents are capable of independently designing viable and effective roof truss structures, indicating that this technique enables agents to acquire useful design methodologies [34].


  1. Gramblička, S.; Kohár, R.; Madaj, R. Construction design automatically adjustable mechanism for crane forks. In Proceedings of the 58th International Conference of Machine Design Departments (ICMD 2017), Praha-Suchdol, Czech Republic, 6–8 September 2017; pp. 100–103.
  2. Weis, P.; Kučera, Ľ.; Pecháč, P.; Močilan, M. Modal analysis of gearbox housing with applied load. Procedia Eng. 2017, 192, 953–958.
  3. Cunningham, J.D.; Simpson, T.W.; Tucker, C.S. An investigation of surrogate models for efficient performance-based decoding of 3D point clouds. J. Mech. Des. 2019, 141, 121401.
  4. Raina, A.; McComb, C.; Cagan, J. Learning to design from humans: Imitating human designers through deep learning. J. Mech. Des. 2019, 141, 111102.
  5. Wu, J.; Zhang, C.; Xue, T.; Freeman, B.; Tenenbaum, J. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Adv. Neural Inf. Process. Syst. 2016, 29, 9.
  6. Kučera, Ľ.; Gajdošík, T. The vibrodiagnostics of gears. In Modern Methods of Construction Design: Proceedings of ICMD 2013; Springer International Publishing: Cham, Switzerland, 2014; pp. 113–118.
  7. Stump, G.M.; Miller, S.W.; Yukish, M.A.; Simpson, T.W.; Tucker, C. Spatial grammar-based recurrent neural network for design form and behavior optimization. J. Mech. Des. 2019, 141, 124501.
  8. Majchrak, M.; Kohar, R.; Lukac, M.; Skyba, R. The process of creating a computational 3d model of a harmonic transmission. MM Sci. J 2020, 13, 3926–3931.
  9. Xiong, Y.; Duong, P.L.T.; Wang, D.; Park, S.I.; Ge, Q.; Raghavan, N.; Rosen, D.W. Data-driven design space exploration and exploitation for design for additive manufacturing. J. Mech. Des. 2019, 141, 101101.
  10. Majchrák, M.; Kohár, R.; Kajan, J.; Skyba, R. 3d meshing methods of ball-rolling bearings. Transp. Res. Procedia 2019, 40, 784–791.
  11. Yoo, S.; Lee, S.; Kim, S.; Hwang, K.H.; Park, J.H.; Kang, N. Integrating deep learning into CAD/CAE system: Generative design and evaluation of 3D conceptual whee. Struct. Multidiscip. Optim. 2021, 4, 2725–2747.
  12. Galbavý, M.; Pitoňák, J.; Kučera, L. Powershift differential transmission with three flows of power for hybrid vehicles. In Modern Methods of Construction Design: Proceedings of ICMD 2013; Springer International Publishing: Cham, Switzerland, 2014; pp. 27–33.
  13. Vardhan, H.; Volgyesi, P.; Sztipanovits, J. Machine learning assisted propeller design. In Proceedings of the ACM/IEEE 12th International Conference on Cyber-Physical Systems, Nashville, TN, USA, 19–21 May 2021; pp. 227–228.
  14. Drbúl, M.; Martikáň, P.; Bronček, J.; Litvaj, I.; Svobodová, J. Analysis of roughness profile on curved surfaces. MATEC Web Conf. 2018, 224, 01024.
  15. Zheng, H.; Moosavi, V.; Akbarzadeh, M. Machine learning assisted evaluations in structural design and construction. Autom. Constr. 2020, 119, 103–346.
  16. Kim, S.; Chi, H.G.; Hu, X.; Huang, Q.; Ramani, K. A large-scale annotated mechanical components benchmark for classification and retrieval tasks with deep neural networks. In Proceedings of the Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, 23–28 August 2020; pp. 175–191.
  17. Exploring Your Design Information with Machine Learning Using the Design Graph. Available online: (accessed on 5 February 2023).
  18. Kucera, L.; Gajdac, I.; Kamas, P. Computing and design of electric vehicles. In Modern Methods of Construction Design: Proceedings of ICMD 2013; Springer International Publishing: Cham, Switzerland, 2014; pp. 105–111.
  19. Vurtur Badarinath, P.; Chierichetti, M.; Davoudi Kakhki, F. A machine learning approach as a surrogate for a finite element analysis: Status of research and application to one dimensional systems. Sensors 2021, 5, 1654.
  20. Shah, K.; Chandnani, R.; Mavinkurve, U.; Raykar, N. Application of Machine Learning for Design-by-Analysis of Pressure Equipment. In Proceedings of the 2019 International Conference on Nascent Technologies in Engineering (ICNTE), Navi Mumbai, India, 4–5 January 2019; pp. 1–6.
  21. Granados-Ortiz, F.J.; Ortega-Casanova, J. Machine learning-aided design optimization of a mechanical micromixer. J. Phys. Fluids 2021, 6, 063604.
  22. Groensfelder, T.; Giebeler, F.; Geupel, M.; Schneider, D.; Jaeger, R. Application of machine learning procedures for mechanical system modeling: Capabilities and caveats to prediction-accuracy. Adv. Model. Simul. Eng. Sci. 2020, 7, 26.
  23. Lynch, M.E.; Sarkar, S.; Maute, K. Machine learning to aid tuning of numerical parameters in topology optimization. J. Mech. Des. 2019, 141, 114502.
  24. Jankejech, P.; Fabian, P.; Broncek, J.; Shalapko, Y. Influence of tempering on mechanical properties of induction bents below 540 C. Acta Mech. Et Autom. 2016, 2, 81–86.
  25. Xu, Y.; Zhang, M.; Zheng, B. Design of cold-formed stainless steel circular hollow section columns using machine learning methods. Structures 2021, 33, 2755–2770.
  26. Sarothi, S.Z.; Ahmed, K.S.; Khan, N.I.; Ahmed, A.; Nehdi, M.L. Predicting bearing capacity of double shear bolted connections using machine learning. J. Mech. Des. 2022, 251, 113–497.
  27. Bécsi, T.; Szabó, Á.; Kővári, B.; Aradi, S.; Gáspár, P. Reinforcement learning based control design for a floating piston pneumatic gearbox actuator. IEEE Access 2020, 8, 147295–147312.
  28. Gramblička, S.; Kohár, R.; Majchrák, M.; Vrabec, M. Contact Analysis of Selected Toothed Contact of the Two-Stage Front Gearbox. In Current Methods of Construction Design: Proceedings of the ICMD 2018; Springer International Publishing: Cham, Switzerland, 2020; pp. 263–269.
  29. Suryadi, D.; Kim, H.M. A data-driven methodology to construct customer choice sets using online data and customer reviews. J. Mech. Des. 2019, 141, 111103.
  30. El Dehaibi, N.; Goodman, N.D.; MacDonald, E.F. Extracting customer perceptions of product sustainability from online reviews. J. Mech. Des. 2019, 141, 121103.
  31. Suryadi, D.; Kim, H.M. A data-driven approach to product usage context identification from online customer reviews. J. Mech. Des. 2019, 141, 121104.
  32. He, Y.; Camburn, B.; Liu, H.; Luo, J.; Yang, M.; Wood, K. Mining and representing the concept space of existing ideas for directed ideation. J. Mech. Des. 2019, 141, 121101.
  33. Raina, A.; Cagan, J.; McComb, C. Transferring design strategies from human to computer and across design problems. J. Mech. Des. 2019, 141, 114501.
  34. Raina, A.; McComb, C.; Cagan, J. Learning to design from humans: Imitating human designers through deep learning. J. Mech. Des. 2019, 141, 111102.
Contributors MDPI registered users' name will be linked to their SciProfiles pages. To register with us, please refer to : , , , , ,
View Times: 914
Revisions: 3 times (View History)
Update Date: 15 Jun 2023
Video Production Service