Design Automation Research Articles

TyBox: An Automatic Design and Code Generation Toolbox for TinyML Incremental On-Device Learning

Incremental on-device learning is one of the most relevant and interesting challenges in the field of Tiny Machine Learning (TinyML). Indeed, differently from traditional TinyML solutions where the training is typically carried out on the Cloud and inference only occurs on the tiny devices (e.g., embedded systems or Internet-of-Things units), incremental on-device TinyML allows both the inference and the training of TinyML models directly on tiny devices. This ability paves the way for TinyML-enabled intelligent devices that can learn directly on the field and adapt to evolving environments, different working conditions, or specific users. The literature in this field is quite limited with very few solutions focusing only on the incremental fine-tuning of machine learning models, whereas a general solution encompassing algorithms and code generation for incremental on-device TinyML is still perceived as missing. The aim of this article is to introduce, to the best of our knowledge for the first time in the literature, a toolbox called TyBox for the automatic design and code generation of incremental on-device TinyML classification models. In more detail, starting from a “static” TinyML model, TyBox is able to (i) automatically design the “incremental” on-device version of the TinyML model that has been suitably designed to take into account the technological constraint on the RAM memory of the target tiny device, and (ii) autonomously provide the C++ codes and libraries to support the inference and learning of the incremental on-device TinyML model directly on the tiny devices. TyBox has been extensively compared with a state-of-the-art incremental learning solution for TinyML and tested on an off-the-shelf tiny device (i.e., the Arduino Nano 33 BLE) in three relevant TinyML application tasks and scenarios: binary image classification, multi-class image classification, and ultra-wide-band human activity recognition. In addition, TyBox is released to the scientific community as a public repository.

Automatic evolutionary design of quantum rule-based systems and applications to quantum reinforcement learning

Explainable artificial intelligence is a research topic whose relevance has increased in recent years, especially with the advent of large machine learning models. However, very few attempts have been proposed to improve interpretability in the case of quantum artificial intelligence, and many existing quantum machine learning models in the literature can be considered almost as black boxes. In this article, we argue that an appropriate semantic interpretation of a given quantum circuit that solves a problem can be of interest to the user not only to certify the correct behavior of the learned model, but also to obtain a deeper insight into the problem at hand and its solution. We focus on decision-making problems that can be formulated as classification tasks and propose a method for learning quantum rule-based systems to solve them using evolutionary optimization algorithms. The approach is tested to learn rules that solve control and decision-making tasks in reinforcement learning environments, to provide interpretable agent policies that help to understand the internal dynamics of an unknown environment. Our results conclude that the learned policies are not only highly explainable, but can also help detect non-relevant features of problems and produce a minimal set of rules.

Automatic Design of Energy-Efficient Dispatching Rules for Multi-Objective Dynamic Flexible Job Shop Scheduling Based on Dual Feature Weight Sets

Considering the requirements of the actual production scheduling process, the utilization of the genetic programming hyper-heuristic (GPHH) approach to automatically design dispatching rules (DRs) has recently emerged as a popular optimization approach. However, the decision objects and decision environments for routing and sequencing decisions are different in the dynamic flexible job shop scheduling problem (DFJSSP), leading to different required feature information. Traditional algorithms that allow these two types of scheduling decisions to share one common feature set are not conducive to the further optimization of the evolved DRs, but instead introduce redundant and unnecessary search attempts for algorithm optimization. To address this, some related studies have focused on customizing the feature sets for both routing and sequencing decisions through feature selection when solving single-objective problems. While being effective in reducing the search space, the selected feature sets also diminish the diversity of the obtained DRs, ultimately impacting the optimization performance. Consequently, this paper proposes an improved GPHH with dual feature weight sets for the multi-objective energy-efficient DFJSSP, which includes two novel feature weight measures and one novel hybrid population adjustment strategy. Instead of selecting suitable features, the proposed algorithm assigns appropriate weights to the features based on their multi-objective contribution, which could provide directional guidance to the GPHH while ensuring the search space. Experimental results demonstrate that, compared to existing studies, the proposed algorithm can significantly enhance the optimization performance and interpretability of energy-efficient DRs.

Enhancing Autonomous Underwater Vehicle Decision Making through Intelligent Task Planning and Behavior Tree Optimization

The expansion of underwater scenarios and missions highlights the crucial need for autonomous underwater vehicles (AUVs) to make informed decisions. Therefore, developing an efficient decision-making framework is vital to enhance productivity in executing complex tasks within tight time constraints. This paper delves into task planning and reconstruction within the AUV control decision system to enable intelligent completion of intricate underwater tasks. Behavior trees (BTs) offer a structured approach to organizing the switching structure of a hybrid dynamical system (HDS), originally introduced in the computer game programming community. In this research, an intelligent search algorithm, MCTS-QPSO (Monte Carlo tree search and quantum particle swarm optimization), is proposed to bolster the AUV’s capacity in planning complex task decision control systems. This algorithm tackles the issue of the time-consuming manual design of control systems by effectively integrating BTs. By assessing a predefined set of subtasks and actions in tandem with the complex task scenario, a reward function is formulated for MCTS to pinpoint the optimal subtree set. The QPSO algorithm is then leveraged for subtree integration, treating it as an optimal path search problem from the root node to the leaf node. This process optimizes the search subtree, thereby enhancing the robustness and security of the control architecture. To expedite search speed and algorithm convergence, this paper recommends reducing the search space by pre-grouping conditions and states within the behavior tree. The efficacy and superiority of the proposed algorithm are validated through security and timeliness evaluations of the BT, along with comparisons with other algorithms for automatic AUV decision control behavior tree design. Ultimately, the effectiveness and superiority of the proposed algorithm are corroborated through simulations on a multi-AUV complex task platform, showcasing its practical applicability and efficiency in real-world underwater scenarios.

SWAN: A multihead autoregressive attention model for solar wind speed forecasting

High speed streams of solar wind can impact Earth’s magnetic environment. Current state of the art geomagnetic effect forecasting methods can predict perturbations up to two hours in advance by using solar wind measurements made in the L1 Lagrange point. By forecasting these solar wind parameters, it should be possible to extend the time range for which said methods can be used.We present Solar Wind Attention Network (SWAN), an attention-based autoregressive model for the forecasting of daily average bulk speed values. We apply a novel automatic design method to create its architecture by developing a pipeline of hyperparameter optimization techniques as a proof of concept. Usage of this pipeline leads to a robust, reliable architecture which we empirically verify the convergence of in a variety of problem configurations regarding solar wind speed forecasting.An improvement in terms of root mean squared error (RMSE) is found when compared to WindNet and convolutional neural networks (CNNs) expanding it, even when SWAN does not incorporate images of the solar disk into the input. SWAN’s RMSE of 68.79 ± 1.05 km s−1 with a lead time of three days compares to RMSE values in the order of 80.28 ± 3.05, 76.30 ± 1.87 and 71.93 ± 0.40 km s−1. From these results, we discuss the possibility of incorporating both autoregressive and image information as input to enhance predictive capability of newly developed models. Furthermore, we analyze the resulting model’s behavior to characterize its main error sources, and find its performance to be at its best in scenarios not involving newly formed coronal holes or coronal mass ejections. It is found that this behavior matches well with a priori expectations derived from the underlying physics. Finally, we briefly discuss the need for an in-depth study of different evaluation metrics to provide a more thorough understanding of modeling results on the field.

Automated intelligent design of modified uni-traveling carrier photodectors

This paper introduces an automatic intelligent design method for the modified uni-traveling carrier photodetector (MUTC-PD). The conventional photodetector design process often relies on the numerical solution of complex nonlinear partial differential equations to simulate and optimize device performance, which is not only computationally intensive but also inefficient. To overcome this challenge, we apply the charge control principle to calculate the photodetector bandwidth, which improves the computational speed by a factor of approximately 1800 compared to the numerical solution of nonlinear partial differential equations. To further optimize the structure of the photodetector, we incorporate the Velocity Varying Climbing Particle Swarm Optimization (VVCPSO) algorithm. This is an improved algorithm based on the traditional particle swarm algorithm, which is able to quickly find the optimal solution in a complex parameter space. By applying the VVCPSO algorithm, we successfully fine-tuned the photodetector structure and obtained structural parameters with optimal performance. Our thorough verification process confirms that the proposed method is consistent with the results of ATLAS simulation software. Automated design has resulted in a high-performance MUTC-PD with a responsivity of 0.52A/W and a bandwidth of 60 GHz (@-3 V) at a mesa diameter of 16µm. Compared to the pre-optimized device, the bandwidth is increased to three times the original. By reducing the mesa diameter to 4µm, the bandwidth can be further increased to 82 GHz (@-3 V). The proposed method's calculation speed is fast enough, enabling extensive parameter studies to optimize device performance.

DARTS-based morphological neural network design and application in bearing fault diagnosis

Morphological filters are widely used in data preprocessing, whereas the existing morphological filters are not optimal in terms of performance, since they are designed by combining the basic morphological operators together artificially. Hence, we proposed an automatic morphological neural network design method based on differentiable architecture search (DARTS), which is a neural architecture search (NAS) method. Firstly, the mechanism is revealed that the morphological dilation with flat structural elements behaves the same as Maxpooling without down-sampling. Secondly, morphological dilation and erosion neural network layers were defined. Thirdly, using DARTS and the search space with morphological neural network layers, end-to-end morphological neural network was implemented. Finally, the morphological neural network was successfully applied in bearing fault diagnosis. Experiments on the three datasets show that the proposed method improved the feature extraction capabilities of neural networks, then achieved end-to-end bearing fault diagnosis, and the diagnosis accuracy was improved by more than 2.0%.

SpikeExplorer: Hardware-Oriented Design Space Exploration for Spiking Neural Networks on FPGA

One of today’s main concerns is to bring artificial intelligence capabilities to embedded systems for edge applications. The hardware resources and power consumption required by state-of-the-art models are incompatible with the constrained environments observed in edge systems, such as IoT nodes and wearable devices. Spiking Neural Networks (SNNs) can represent a solution in this sense: inspired by neuroscience, they reach unparalleled power and resource efficiency when run on dedicated hardware accelerators. However, when designing such accelerators, the amount of choices that can be taken is huge. This paper presents SpikExplorer, a modular and flexible Python tool for hardware-oriented Automatic Design Space Exploration to automate the configuration of FPGA accelerators for SNNs. SpikExplorer enables hardware-centric multiobjective optimization, supporting target factors such as accuracy, area, latency, power, and various combinations during the exploration process. The tool searches the optimal network architecture, neuron model, and internal and training parameters leveraging Bayesian optimization, trying to reach the desired constraints imposed by the user. It allows for a straightforward network configuration, providing the full set of explored points for the user to pick the trade-off that best fits their needs. The potential of SpikExplorer is showcased using three benchmark datasets. It reaches 95.8% accuracy on the MNIST dataset, with a power consumption of 180 mW/image and a latency of 0.12 ms/image, making it a powerful tool for automatically optimizing SNNs.

Efficient complete denture metal base design via a dental feature-driven segmentation network

Background and objectiveComplete denture is a common restorative treatment in dental patients and the design of the core components (major connector and retentive mesh) of complete denture metal base (CDMB) is the basis of successful restoration. However, the automated design process of CDMB has become a challenging task primarily due to the complexity of manual interaction, low personalization, and low design accuracy. MethodsTo solve the existing problems, we develop a computer-aided Segmentation Network-driven CDMB design framework, called CDMB-SegNet, to automatically generate personalized digital design boundaries for complete dentures of edentulous patients. Specifically, CDMB-SegNet consists of a novel upright-orientation adjustment module (UO-AM), a dental feature-driven segmentation network, and a specific boundary-optimization design module (BO-DM). UO-AM automatically identifies key points for locating spatial attitude of the three-dimensional dental model with arbitrary posture, while BO-DM can result in smoother and more personalized designs for complete denture. In addition, to achieve efficient and accurate feature extraction and segmentation of 3D edentulous models with irregular gingival tissues, the light-weight backbone network is also incorporated into CDMB-SegNet. ResultsExperimental results on a large clinical dataset showed that CDMB-SegNet can achieve superior performance over the state-of-the-art methods. Quantitative evaluation (major connector/retentive mesh) showed improved Accuracy (98.54 ± 0.58 %/97.73 ± 0.92 %) and IoU (87.42 ± 5.48 %/70.42 ± 7.95 %), and reduced Maximum Symmetric Surface Distance (4.54 ± 2.06 mm/4.62 ± 1.68 mm), Average Symmetric Surface Distance (1.45 ± 0.63mm/1.28 ± 0.54 mm), Roughness Rate (6.17 ± 1.40 %/6.80 ± 1.23 %) and Vertices Number (23.22 ± 1.85/43.15 ± 2.72). Moreover, CDMB-SegNet shortened the overall design time to around 4 min, which is one tenth of the comparison methods. ConclusionsCDMB-SegNet is the first intelligent neural network for automatic CDMB design driven by oral big data and dental features. The designed CDMB is able to couple with patient's personalized dental anatomical morphology, providing higher clinical applicability compared with the state-of-the-art methods.

The Genesis of AIbyAI Integrated Circuit: Where AI Creates AI

The typical Integrated Circuit (IC) development process commences with formulating specifications in natural language and subsequently proceeds to Register Transfer Level (RTL) implementation. RTL code is traditionally generated through manual efforts, using Hardware Description Languages (HDL) such as VHDL or Verilog. High-Level Synthesis (HLS), on the other hand, converts programming languages to HDL; these methods aim to streamline the engineering process, minimizing human effort and errors. Currently, Electronic Design Automation (EDA) algorithms have been improved with the use of AI, with new advancements in commercial (such as ChatGPT, Bard, among others) Large Language Models (LLM) and open-source tools presenting an opportunity to automate the chip design process. This paper centers on the creation of AIbyAI, a Convolutional Neural Network (CNN) IC entirely developed by an LLM (ChatGPT-4), and its manufacturing with the first fabricable open-source Process Design Kit (PDK), SKY130A. The challenges, opportunities, advantages, disadvantages, conversation flow, and workflow involved in CNN IC development are presented in this work, culminating in the manufacturing process of AIbyAI using a 130 nm technology, marking a groundbreaking achievement as possibly the world’s first CNN entirely written by AI for its IC manufacturing with a free PDK, being a benchmark for systems that can be generated today with LLMs.

Design and Test of Offset Quadrature Phase-Shift Keying Modulator with GF180MCU Open Source Process Design Kit

This article explores the evolution of integrated circuits (ICs), highlighting the fundamental role of open source Electronic Design Automation (EDA) tools in their development. It describes the IC’s design flow, differentiating between Front-end and Back-end design stages, and details the process of implementing the digital stage in offset quadrature phase-shift keying (OQPSK) modulation in an IC, including its hardware description language (HDL), the implementation test in the field-programmable gate array (FPGA), and the physical layout using the first manufactured open source process design kits (PDKs) in Global Foundries’ 180 nm, as well as the use of OpenLane and Caravel. To conclude, the results of the physical tests obtained from the digital modulation are presented, as well as the performance of the raised cosine shaping filter.

A Generative Deep Learning Approach for Improving the Mechanical Performance of Structural Components

This study aimed to improve the mechanical properties of 3D concept designs by combining the design capability of a generative adversarial network with finite element analysis. This approach offers an innovative perspective on the conditioning of generative models while improving design properties and automation. A new design and evaluation framework has been developed for GAN models to generate 3D models with improved mechanical properties. The framework is an iterative process that includes dataset generation, GAN training, and finite element analysis. A “joint” component used in the aerospace industry is considered to demonstrate the proposed method’s effectiveness. Over six iterations, an increase of 20% is recorded in the average safety factor of the designs, and the variety of designs produced is narrowed in the desired direction. These findings suggest that the direct generation of structural components with generative models can expand the potential of deep learning in engineering design. Another innovative aspect of this study is that it provides a new option for the conditioning of data-dependent generative design models.

Automatic design of interpretable control laws through parametrized Genetic Programming with adjoint state method gradient evaluation

This work investigates the application of a Local Search (LS) enhanced Genetic Programming (GP) algorithm to the control scheme’s design task. The combination of LS and GP aims to produce an interpretable control law as similar as possible to the optimal control scheme reference. Inclusive Genetic Programming (IGP), a GP heuristic capable of promoting and maintaining the population diversity, is chosen as the GP algorithm since it proved successful on the considered task. IGP is enhanced with the Operators Gradient Descent (OPGD) approach, which consists of embedding learnable parameters into the GP individuals. These parameters are optimized during and after the evolutionary process. Moreover, the OPGD approach is combined with the adjoint state method to evaluate the gradient of the objective function. The original OPGD was formulated by relying on the backpropagation technique for the gradient’s evaluation, which is impractical in an optimization problem involving a dynamical system because of scalability and numerical errors. On the other hand, the adjoint method allows for overcoming this issue. Two experiments are formulated to test the proposed approach, named Operator Gradient Descent - Inclusive Genetic Programming (OPGD-IGP): the design of a Proportional–Derivative (PD) control law for a harmonic oscillator and the design of a Linear Quadratic Regulator (LQR) control law for an inverted pendulum on a cart. OPGD-IGP proved successful in both experiments, being capable of autonomously designing an interpretable control law similar to the optimal ones, both in terms of shape and control gains.

Automatic design of W-operators using membership functions: a case study in brain MRI segmentation

AbstractA W-operator is an image transformation that is locally defined inside a window W, invariant to translations. The automatic design of the W-operators consists of the design of functions, whose domain is a set of patterns or vectors obtained by translating a window through training images and the output of each vector is a class or label. The main difficulty to consider when designing W-operators is the generalization problem that occurs due to lack of training images. In this work, we propose the use of membership functions to solve the generalization problem in gray level images. Membership functions are defined from the training images to model regions that are often inaccurate due to ambiguous gray levels in the images. This proposal was applied to brain magnetic resonance image segmentation to test its performance in a field of interest in biomedical images. The experiments were carried out with different numbers of training and test images, windows sizes of $$3\times 3$$ 3 × 3 , $$5\times 5$$ 5 × 5 , $$7\times 7$$ 7 × 7 , $$11\times 11$$ 11 × 11 , and $$15\times 15$$ 15 × 15 , and images with noise levels at 0, 1, 3, 5, 7, and 9$$\%$$ % . To calculate the performance of each designed W-operator, the classification error, sensitivity, and specificity were used. From the experimental results, it was concluded that the best performance is achieved with a window of size $$3\times 3$$ 3 × 3 . In images with noise levels from 1 to 5$$\%$$ % , the classification error is less than 4$$\%$$ % and the sensitivity and specificity are greater than 94 and 98$$\%$$ % , respectively.

Oscillatory Neural Network-Based Ising Machine Using 2D Memristors.

Neural networks are increasingly used to solve optimization problems in various fields, including operations research, design automation, and gene sequencing. However, these networks face challenges due to the nondeterministic polynomial time (NP)-hard issue, which results in exponentially increasing computational complexity as the problem size grows. Conventional digital hardware struggles with the von Neumann bottleneck, the slowdown of Moore's law, and the complexity arising from heterogeneous system design. Two-dimensional (2D) memristors offer a potential solution to these hardware challenges, with their in-memory computing, decent scalability, and rich dynamic behaviors. In this study, we explore the use of nonvolatile 2D memristors to emulate synapses in a discrete-time Hopfield neural network, enabling the network to solve continuous optimization problems, like finding the minimum value of a quadratic polynomial, and tackle combinatorial optimization problems like Max-Cut. Additionally, we coupled volatile memristor-based oscillators with nonvolatile memristor synapses to create an oscillatory neural network-based Ising machine, a continuous-time analog dynamic system capable of solving combinatorial optimization problems including Max-Cut and map coloring through phase synchronization. Our findings demonstrate that 2D memristors have the potential to significantly enhance the efficiency, compactness, and homogeneity of integrated Ising machines, which is useful for future advances in neural networks for optimization problems.

Analysis of automatic design systems

The process of designing and developing a control program begins after the previous stage of the product life cycle has been completed and it begins with the work of the design bureau. It develops new design solutions, and the technical specification for the required product is agreed with the customer. Based on this, a product drawing with technical requirements is created, which is usually transmitted to the department (technologists-programmers), which is directly involved in the design and development of process control for the product. At this stage, as a rule, the technologist consults with the developer of the drawing for the required product to obtain complete information about the product in order to eliminate any errors. Such cases are not excluded when, from this moment on, the drawing is returned for revision to the department of the design bureau due to the inability to sell the product in metal due to lack of industrial capacity or poor design solutions that make the description of the manufacturing technology impossible. After receiving the terms of reference, the next step is the development of the technological process. The technological process must be compiled technically competently in accordance with the internal standards of the enterprise (everyone may have their own, but the structure of the compilation is the same) and be informative enough at least to eliminate the high probability of defects during manufacture. As statistics show, in many enterprises, the overwhelming percentage of defects appears due to the negligent attitude of operators of CNC machines due to deviations from the technological process.

A novel framework for set-based steel connection design automation

Existing computational design tools for steel connection design predominantly employ a point-based design (PBD) approach, which requires iterative re-evaluation whenever there are changes in design specifications. This paper introduces a new framework that adopts a set-based design (SBD) approach, aiming to substantially reduce the iteration and time cost associated with steel connection design and rework. The framework integrates a component-set connection design model with a database storage and query-based data retrieval method. The first method enables the flexible and efficient generation of a large connection design space from all possible component combinations, and automated identification of valid connection configurations within it. The latter method allows for automated design space refinement from preference-based evaluation of connection design efficiency and high-speed comparison and selection of optimal connection designs. To evaluate the relative performance of SBD and PBD approaches, the framework is applied to a steel floor system design case study with 62 connections. Results showed that the SBD approach achieved near-instantaneous connection design automation, with a total execution time of fewer than 55 milliseconds, making it over 10 times faster than the corresponding PBD approach.

Accelerating Understanding of Human Response to Automation Failure

Firstly, I comment on the lack of support for the predictions of the lumberjack model to professionally qualified operators in high-fidelity work simulations (Jamieson & Skraaning, 2020a). I highlight the advantages that Bayesian statistics provide for qualifying the degree of evidence for the null hypotheses, issues concerning situation awareness measurement, and the alternative techniques available to study experts. Secondly, I comment on the innovative taxonomy of automation failure presented by Skraaning and Jamieson (2024), pointing out some issues with overlapping definitions and lack of cause-effect relationships. I then discuss the substantial opportunity this taxonomy presents to guide future research, such as the design of transparent automation. To conclude, I identify some other key problems regarding how we currently study human-automation teaming (e.g. presenting randomized automation failure unlinked to task context) and invite discussion from the research community on the relevance of computational modelling to this field of research.

Automatic design of gene regulatory mechanisms for spatial pattern formation

Gene regulatory mechanisms (GRMs) control the formation of spatial and temporal expression patterns that can serve as regulatory signals for the development of complex shapes. Synthetic developmental biology aims to engineer such genetic circuits for understanding and producing desired multicellular spatial patterns. However, designing synthetic GRMs for complex, multi-dimensional spatial patterns is a current challenge due to the nonlinear interactions and feedback loops in genetic circuits. Here we present a methodology to automatically design GRMs that can produce any given two-dimensional spatial pattern. The proposed approach uses two orthogonal morphogen gradients acting as positional information signals in a multicellular tissue area or culture, which constitutes a continuous field of engineered cells implementing the same designed GRM. To efficiently design both the circuit network and the interaction mechanisms—including the number of genes necessary for the formation of the target spatial pattern—we developed an automated algorithm based on high-performance evolutionary computation. The tolerance of the algorithm can be configured to design GRMs that are either simple to produce approximate patterns or complex to produce precise patterns. We demonstrate the approach by automatically designing GRMs that can produce a diverse set of synthetic spatial expression patterns by interpreting just two orthogonal morphogen gradients. The proposed framework offers a versatile approach to systematically design and discover complex genetic circuits producing spatial patterns.

METHOD OF GENERATIVE-ADVERSARIAL NETWORKS SEARCHING ARCHITECTURES FOR BIOMEDICAL IMAGES SYNTHESIS

Context. The article examines the problem of automatic design of architectures of generative-adversarial networks. Generativeadversarial networks are used for image synthesis. This is especially true for the synthesis of biomedical images – cytological and histological, which are used to make a diagnosis in oncology. The synthesized images are used to train convolutional neural networks. Convolutional neural networks are currently among the most accurate classifiers of biomedical images.
 Objective. The aim of the work is to develop an automatic method for searching for architectures of generative-adversarial networks based on a genetic algorithm.
 Method. The developed method consists of the stage of searching for the architecture of the generator with a fixed discriminator and the stage of searching for the architecture of the discriminator with the best generator.
 At the first stage, a fixed discriminator architecture is defined and a generator is searched for. Accordingly, after the first step, the architecture of the best generator is obtained, i.e. the model with the lowest FID value.
 At the second stage, the best generator architecture was used and a search for the discriminator architecture was carried out. At each cycle of the optimization algorithm, a population of discriminators is created. After the second step, the architecture of the generative-adversarial network is obtained.
 Results. Cytological images of breast cancer on the Zenodo platform were used to conduct the experiments. As a result of the study, an automatic method for searching for architectures of generatively adversarial networks has been developed. On the basis of computer experiments, the architecture of a generative adversarial network for the synthesis of cytological images was obtained. The total time of the experiment was ~39.5 GPU hours. As a result, 16,000 images were synthesized (4000 for each class). To assess the quality of synthesized images, the FID metric was used.The results of the experiments showed that the developed architecture is the best. The network’s FID value is 3.39. This result is the best compared to well-known generative adversarial networks.
 Conclusions. The article develops a method for searching for architectures of generative-adversarial networks for the problems of synthesis of biomedical images. In addition, a software module for the synthesis of biomedical images has been developed, which can be used to train CNN.