Generative Design Based on Reverse and Forward Modeling Machine Learning

Abstract

A network for machine learning provides generative design. Instead of mimicking the typical human design process, machine training is performed on an inverse model to generate designs according to requirements. Machine training is performed on a simulation model to recover performance with respect to the requirements of the generated designs. These two machine-trained models are used in an optimization that creates other designs based on the output of the inverse model and tests those designs with the simulation model. Using the machine-trained models in this loop to explore a variety of different designs reduces the time for exploration and thus can bring more optimized designs to design engineers or enable them to better initiate designs.

Description

Background Art

[0001] This embodiment relates to the design of objects such as high-efficiency low-emission gas turbines. Such designs consider attributes such as temperature, turbine shaft power, turbine exhaust pressure, and other characteristics. In the turbine use case, given the limitations for a given design, the efficiency will be maximized while keeping the exhaust temperature below a maximum value and maintaining the ambient temperature and turbine shaft power fixed.

[0002] Design engineers use these parameters, constraints, and evaluation criteria (e.g., the shaft power must be within 0.05% of the design standard and the total error must be within 5% of the design parameters) to design efficient gas turbines and engines. Design engineers spend a lot of time using a large number of time-consuming simulation-based tools to generate new design iterations. For complex turbine systems, it may take hours to verify using simulation software, thus slowing down the design-evaluation-redesign cycle. As a result, due to the rather time-consuming design creation and analysis process, design engineers may not be able to explore the complete design space. It may not be possible to produce the most efficient optimal design. Summary of the Invention

[0003] As an introduction, the preferred embodiments described below include methods, systems, instructions, and computer-readable media for generative design through a machine learning network. Instead of mimicking the typical human design process, an inverse model machine learning model is trained to generate designs according to requirements. A surrogate model (simulation or machine learning) is trained to estimate the engine performance of a given design. These two trained machine learning models are used to optimize the design. Other designs are created according to the inverse model to output designs, and the surrogate model is used to test those designs. Using machine learning models in this loop to explore a variety of different designs reduces the exploration time and thus can bring more optimized designs to design engineers or enable them to better start the design.

[0004] In a first aspect, a method for generative design through an artificial intelligence processor is provided. One or more designs of a first object (target) to be designed are generated by the artificial intelligence processor. The generation is performed by a first (e.g., inverse) machine learning network in response to design requirements, constraints, and goals of the first object to be designed. The artificial intelligence processor simulates the operation of each of the one or more designs through a second machine learning network (e.g., surrogate model). The simulation provides a seed design based on design requirements, constraints, and / or goals. One or more designs are perturbed, and the simulation is repeated for the perturbed one or more designs. An error of the perturbed one or more designs is determined, which is the error between a second value from the simulation and a first value (e.g., design output and those goals). Based on the error, at least one of the perturbed or generated one or more designs is selected. At least one of the selected perturbed or generated one or more designs is stored.

[0005] In a second aspect, a system for machine learning-based design is provided. A processor is configured by instructions stored in a memory. When executed by the processor, the instructions are for: inverse modeling design parameters according to a design specification, the inverse modeling using a first neural network trained by a first machine; optimizing the design parameters based on simulating the design using the design parameters from the inverse modeling, the simulation using a second neural network trained as a forward model or surrogate model by the first machine or a second machine; and outputting design parameters for the design.

[0006] In a third aspect, a method for a machine learning design system is provided. A machine trains a first neural network (e.g., generative model or inverse model) as a generative model to inverse model to generate a design (e.g., model a set of designs according to functional requirements) based on a first value of design requirements. A machine or another machine trains a second neural network as a prediction model or surrogate model to predict a value of performance based on design values. The design system is programmed using the second neural network (e.g., a combination of the design and the second neural network) to optimize the design.

[0007] The invention is defined by the following claims, and this section should not be regarded as a limitation of those claims. Other aspects and advantages of the invention are discussed below in connection with the preferred embodiments, and may be claimed independently or in combination subsequently. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The components and the drawings are not necessarily drawn to scale, but rather emphasize the illustration of the principles of the invention. Further, in the drawings, the same reference numerals denote corresponding components in different views.

[0009] Figure 1A flowchart of an embodiment of a method for machine training a design system to generate a design;

[0010] Figure 2 A flowchart of an embodiment of a method for generative design by an artificial intelligence processor;

[0011] Figure 3 Illustrates an example of generating a machine learning design based on requirements and constraints;

[0012] Figure 4 Illustrates an example of a machine learning simulation of performance based on a design; and

[0013] Figure 5 A block diagram of an embodiment of a system for machine learning-based design. Detailed Description

[0014] Machine learning-driven optimization is used for generative design. Machine learning-based generative design generates new designs of objects, such as turbine designs. Machine learning-based generative design follows user-defined performance or design parameters and criteria. In addition, machine learning-based optimization techniques improve the design to meet user-defined functional requirements and achieve certain purposes.

[0015] Using machine learning-based design generation and optimization enables the processor to execute design runs faster. Compared with manual design with programmed simulations, the processor runs better and more efficiently with a machine learning-based surrogate model. The machine learning-based design generation and optimization are automatic after initial training using historical data, only requiring the user to specify the desired attributes of the design and the evaluation criteria for accepting or rejecting the automatic design. Since the design system can autonomously suggest designs that meet new requirements, the machine learning-based autonomous design advisor speeds up the design lifecycle, enabling new designs to be quickly created and evaluated. The use of machine learning can better explore design options for generating designs, enabling the opportunity to explore more design options within a given cycle time or shortening the design cycle time through faster decision-making.

[0016] The object to be designed can be of any type. In the following example, a turbine, such as a low-emission gas turbine, is designed. Other objects that can be generatively designed include suspensions, printed circuit boards, blades (e.g., wind turbine blades), shafts, circuits, transmissions, transformers, impellers, the topologies of any type of component, the structures of any type of component (e.g., wheel or hull designs), routers or blades, or other objects.

[0017] Figure 1An embodiment of a method for machine training of a design system is shown. The training uses deep learning to train both a generator and a simulator or surrogate model of the design to simulate the performance of the generated design. The simulator can be used to optimize the design to more thoroughly explore possible designs.

[0018] The method is implemented by a machine, a database, and / or a programmer. The programmer can program manually or create software to optimize. The machine uses deep learning to create a generative model and a predictive simulator model, thus providing a machine-based rather than a human-based method for design generation and performance simulation. The training data for deep learning (e.g., designs and performance measurements) are stored in a database. A machine such as a processor, a workstation, a server, a controller, or other processors trains a defined neural network architecture based on the training data in the database.

[0019] The actions are performed in the order shown (i.e., top to bottom or numerical order) or in other orders. Additional, different, or fewer actions may be provided. For example, action 16 is not provided. As another example, action 12 is performed using a database of the created training data. In another example, action 12 is divided into two actions, one for training a design generative model and the other for training a simulation model. In yet another example, actions for defining a neural network architecture, activating training, selecting training data, testing a machine learning network, and / or controlling learning are provided.

[0020] In action 10, training data are obtained. For machine training, hundreds, thousands, tens of thousands, or other quantities of samples with ground truth are used. The generative model takes expected performance information (such as operating requirements and constraints) as input and outputs design information (such as settings). This is a reverse approach. The simulation model is a forward predictor and thus receives design information and outputs performance information. The same training data can be used to train both models, where the ground truth of one model is the input sample of the other model and vice versa. In alternative embodiments, different training data are obtained to train different models.

[0021] Training data are collected from sensors, measurements, and / or other records. For implemented designs, performance measurements are obtained from on-site measurements or tests. Design information of the implemented design is obtained from a database or record of the design. For other samples in the training data, physical simulations are used for proposed or previous simulations. The proposed design information and simulation performance provide samples. Design tools may be used.

[0022] Performance information may include user requirements and / or constraints. The required operating specifications of the design and any acceptable ranges (e.g., minimum and / or maximum values) define the parameters of the performance information as constraints. Performance information is the operating characteristics of the design, the size and weight of the design, etc. Requirements may be performance parameters with values to be fixed by the design. Constraints may be performance parameters with values constrained to a maximum and / or minimum value. Performance information may include one or more performance parameters of a target. One or more requirements and / or constraints may be identified as a target. For example, it is required to optimize efficiency (e.g., turbine efficiency). A target is a performance parameter under which the design will operate in the best possible way or better than other designs.

[0023] In an example of a gas turbine, the user inputs values of performance parameters such as ambient temperature, turbine shaft power, turbine exhaust temperature, and inlet flow rate of the burner cooling section. The target is to maximize the performance parameter of turbine efficiency while satisfying the above constraints and / or requirements. Values of other performance parameters can be used as requirements or constraints, such as lower heating value of fuel, external cooler power, (total) pressure of gas turbine exhaust, inlet flow rate of the burner cooling section, burner cooling efficiency, burner cooling power, presence or absence of external cooling, inlet flow rate of ambient air cooling section, increment compared to any quantity, compressor outlet temperature, fuel flow rate, CO 2 exhaust composition, H 2 O exhaust composition, N 2 exhaust composition, Ar exhaust composition, O 2 exhaust composition, SO 2 exhaust composition, compressor inlet flow rate, turbine inlet temperature, compressor outlet pressure, and / or enthalpy at the compressor outlet. Different performance parameters can be used as parameters for requirements, constraints, and / or targets. Depending on the design needs, some may not be used. Other performance parameters can be used for gas turbines or other objects. Any number of performance parameters and corresponding requirements or constraints can be used.

[0024] Design information includes settings for controlled characteristics. Specific choices in the design may result in specific design characteristics. Design parameters are variables controlled by the object to produce performance parameters.

[0025] Design parameters may be the same as or different from performance parameters. In cases where performance is not required or constrained, the parameters can be part of the design.

[0026] In one embodiment of a gas turbine example, the design parameters include gas turbine shaft efficiency, polytropic compressor efficiency, isentropic ISO turbine efficiency, ISO turbine inlet temperature, compressor pressure ratio, SAS equivalent mass flow rate, primary zone temperature, gas turbine exhaust flow rate, turbine cooling section inlet flow rate, compressor inlet flow rate, turbine inlet temperature, compressor outlet pressure, enthalpy at the compressor outlet, burner cooling power, cooled cooling air flow rate, compressor outlet temperature, fuel flow rate, CO 2 exhaust composition, H 2 O exhaust composition, N 2 exhaust composition, Ar exhaust composition, and O 2 exhaust composition. In another embodiment, the design parameters include ambient temperature, polytropic compressor efficiency, isentropic ISO turbine efficiency, ISO turbine inlet temperature, compressor pressure ratio, SAS equivalent mass flow rate, primary zone temperature, gas turbine exhaust flow rate, lower heating value of fuel, burner cooling section inlet flow rate, turbine cooling section inlet flow rate, cooled cooling air flow, presence or absence of external cooling, presence or absence of SiCAT, ambient air cooling section inlet flow rate, and an increment compared to any quantity. Additional, different, or fewer parameters may be used for design.

[0027] In operation 12, the machine trains a first neural network as a generative model to inverse model according to the first values of the design requirements and / or constraints to generate a design. The machine trains the network to receive performance parameter values in the inverse model and output design parameter values. The same or a different machine trains a second neural network as a predictive model to predict the values of the design requirements (e.g., the second neural network acts as a surrogate model to simulate the performance of the design). The machine trains the network to receive design parameter values and output performance parameter values. The machine trains the network to simulate the operation of the object being designed.

[0028] Deep learning is used to train artificial neural networks. Any deep learning method or architecture can be used. Deep learning trains filter kernels, connections, weights, and / or other features that together indicate the output value for a given input. Deep learning can be used to learn the features that can be used to determine the output. Deep learning provides a deep machine learning network that outputs a result given a previously unseen input value. The machine uses training data to learn the output value given an unseen input value.

[0029] The architecture of the network may include convolutional, subsampling (e.g., max or average pooling), fully connected layers, recurrent layers, SoftMax, concatenation, dropout, residual, and / or other types of layers. Any combination of layers may be provided. Any layer arrangement may be used. Skip, feedback, or other connections within or between layers may be used. A hierarchical structure may be employed for learning features or representations or for classification or regression.

[0030] The architecture for deep learning may include a convolutional neural network (CNN) or convolutional layers. A CNN defines one or more layers, where each layer has a filter kernel for convolving with the input of that layer. Other convolutional layers may provide further extraction, such as receiving the output from a previous layer and convolving the output with another filter kernel. Machine learning identifies the filter kernels to extract features indicative of solar radiation.

[0031] The architecture may include one or more dense layers. A dense layer connects various features from a previous layer to the output of that layer. In one embodiment, the dense layer is a fully connected layer. One or more fully connected layers are provided. The dense layers may form a multi - layer perceptron.

[0032] In one embodiment, the architecture of the generative model is a generative adversarial network (GAN) or a mixture density network (Gaussian mixture or kernel density estimator). A GAN includes an encoder and a decoder connected to a bottleneck, generating output information based on the input by increasing and then decreasing the extraction level. A discriminator is provided during training to predict the accuracy of the output of the encoder - decoder. This prediction is fed back to the encoder - decoder during training. Both the discriminator and the encoder - decoder learn during training. The architecture of the forward or prediction model for simulation is a fully connected residual network. Other types of architectures or corresponding networks may be used, such as other generative models or feed - forward neural networks for generative models or other residual networks for prediction or simulation models. Other types of machine learning and corresponding networks may be used instead of neural networks and deep learning, such as support vector machines.

[0033] The network is trained using any optimization (such as Adam, a learning rate of 0.0001, and / or mean squared error loss (e.g., L1)) for any number of epochs (e.g., 300). Other arrangements, layers, units, activation functions, architectures, learning rates, optimizations, loss functions, and / or normalizations may be used. Other architectures may be used.

[0034] Train the recognition kernel, weights, connections, and / or other information of the architecture that correlates the input with the ground truth output. The neural network trained by deep learning is stored in action 14. Once the training is complete, the artificial neural network of machine learning can be applied. Store the network so that the network can be used in applications for specific designs.

[0035] Based on the training, the generative or inverse network can output one or more designs that take into account the given user requirements, constraints, and / or goals. The simulation or prediction network can output one or more performance characteristic values for each design. To explore the design space, other designs are tried. In action 16, a program is created to change the proposed design. Then, the machine learning simulation or prediction network is used to simulate the changed design. The machine learning network can be used to quickly test many designs, so the programming optimizes the design based on the output of the machine learning network. The goals and the satisfaction or differences with respect to the requirements and / or constraints are used to optimize the design, thereby generating one or more designs for the object.

[0036] The programming is manual. The programmer programs the machine to use the machine learning network to provide one or more optimized designs.

[0037] Figure 2 An embodiment of a flowchart of a method for generative design of an artificial intelligence processor is shown. The generative and predictive networks of machine learning are used to design an object. The machine learning network is used to optimize the design.

[0038] The artificial intelligence processor is a processor for applying the machine learning network. The artificial intelligence processor can include massive parallel processing, or can be a general-purpose processor, an image processor, a graphics processing unit, an application-specific integrated circuit, a field-programmable gate array, or other processors capable of using the machine learning network. In one embodiment, a Figure 5 design system is used. The processor executes actions 22 - 26. The user or the processor executes actions 20 and 21. The manufacturing system or component can execute action 27. Alternatively, the memory executes action 27.

[0039] The actions are executed in the order shown (i.e., from top to bottom or numerically) or in other orders. For example, actions 20 and 21 are executed simultaneously or in any order. As another example, action 27 can be performed on all designs after action 23 and / or action 25, and / or action 27 can be performed on the design output after action 22 and before action 23.

[0040] Additional, different, or fewer actions may be provided. For example, action 28 is not performed. As another example, actions 26 and / or 27 are not provided. In yet another example, actions for further design modification and / or simulation are provided, such as in the case where optimization is used to provide an initial design improved by a design engineer.

[0041] In action 20, the requirements and constraints of the current project are input. The user provides the specifications of the project. The processor inputs the values of the performance parameters into the generative model.

[0042] In action 21, one or more goals are identified. The user indicates the performance parameters to be optimized, such as indicating gas turbine efficiency. The processor marks or identifies the goals to be used in the optimization of action 24. The goals can also be requirements or constraints and are thus shown as inputs to action 22. In an alternative embodiment, the goals are inputs to the optimization of action 24.

[0043] In action 22, the processor generates one or more designs of the object to be designed. The design is the value of the design parameters indicating the structure, control, and / or manufacture of the object.

[0044] The generative model provided by the machine learning network generates one or more designs. In response to an input or request of the generative model, a single or multiple different possible designs are generated. The generative model is an inverse model that generates designs based on design requirements rather than generating performance parameter values based on input designs. Given performance characteristics, the inverse model designs multiple sets of design values to generate these performance outputs.

[0045] Figure 3 An example is shown. The values of the operating or performance parameters listed as requirements are input into the machine learning generative network 30. The machine learning network 30 outputs the values of the design parameters. The requirements are represented as output variables to be input into the inverse model, and the design parameters are represented as input variables, reflecting inverse modeling.

[0046] The function of the generative model is to take a set of user requirements or goals and generate a set of designs that meet the goals (i.e., design solutions that work given the requirements and constraints). These design solutions (Design 1, Design 2, … Design n) are generative designs of a gas turbine engine that can meet the goals.

[0047] Any number of designs can be output, such as a fixed number or all possible designs that meet the requirements and constraints. For example, three designs or at least three designs (e.g., 1,500) are output in response to the input. The user requirements and constraints (including the target parameters to be maximized or minimized) are input into the generative model, which outputs multiple designs.

[0048] The input target is used to select some or less than all of the output designs. The target is a requirement, constraint, or other performance parameter to be maximized or minimized. Selection is performed on the output of the machine learning inverse network, but the network may be trained to perform the selection. A subset of output designs can be selected based on the target, such as selecting n output designs as the n designs with the maximum efficiency. In one example, 100 designs are selected as initial seeds for each factor value to be maximized or minimized (e.g., maximum efficiency). Selection can be sampled over a range of multiple target values (e.g., multiple efficiencies), such as selecting designs with efficiency values of 100, 95, 90, 85, 80, 75, 70, 68, 65, 63, 60, 58, 55, 53, and 50. Alternatively, all output designs are used (e.g., full selection).

[0049] The selected designs meet the requirements and constraints. Designs that meet some, most, or nearly meet can be output and selected, rather than strictly meet. The result of the generative model is a design that may meet the target (e.g., constraint and / or requirement), but not necessarily the best design.

[0050] The designs output by the generative model are then used by the optimization module to find more optimized designs. The selected designs are used as initial seeds. Given user input and a seed efficiency value, optimization generates additional designs through the inverse model based on these initial seed outputs.

[0051] In action 23, the processor simulates the operation of each selected design. The operation of one or more designs is simulated. The simulation provides values of one or more performance or operation parameters. The simulation can estimate the parameter values of the requirements and / or constraints. The simulation can estimate one or more values of the target. The simulation can estimate the values of other operation or performance parameters.

[0052] The simulation is provided by a forward or predictive model. A trained machine learning network simulates the values of the output performance parameters. The machine learning model outputs the values of the performance parameters for each input design (e.g., design requirements, constraints, and / or targets). Given input features (i.e., input values of design parameters), the forward model predicts the output features (e.g., acting as a surrogate model for the simulation). The predictive model estimates the operation of the proposed design.

[0053] Figure 4 An example is shown. The values of the design parameters are input into the machine learning residual neural network 40. The machine learning network 40 outputs the values of the performance or operation parameters. The performance parameters are the same as or different from the requirements, constraints, and / or targets. The performance parameters are represented as output variables, and the design parameters are represented as input variables, reflecting forward modeling.

[0054] The design was verified through simulation. Each design (e.g., each initial seed) was input sequentially. The simulation machine learning model output the performance parameter values for each input design. The output of the generative model was a set of design parameters to be verified. The simulation model indicated whether the design met the user requirements, constraints, and / or objectives. The machine learning prediction model did not perform expensive and time-consuming simulations but approximated the simulation results within a fraction of a second.

[0055] Another selection can be made. The values of the performance parameters can be used for selection. The initial seeds are evaluated against a criterion (e.g., an objective such as efficiency) using a forward model. The evaluation against requirements and / or constraints can also be used for selection. Promising seeds (i.e., designs) that meet or are close to the criterion are selected.

[0056] The simulation output is the performance result or simulation run of the initial seeds of the designs from the generative model. The initial design seeds that best meet one or more objectives are the optimized designs.

[0057] In action 24, other designs can be created to further optimize the design. The generative model and the simulation model are used to optimize by maximizing or minimizing the value of a target parameter (such as turbine efficiency). The optimization generates m designs close to the best design. The m designs are created based on the n initial design seeds output by action 22 or a subset of the n initial design seeds (e.g., the seeds with the maximum efficiency). The optimization selects the input values of the design parameters such that the output value of the objective is maximized or minimized. Designs can be selected for optimization based on the satisfaction of requirements and / or constraints.

[0058] In action 25, additional designs are generated based on the initial design seeds. The designs are perturbed. The perturbation generates the design output of the generative model or a selected subset of the design output. Previously perturbed designs may be perturbed again.

[0059] The design is perturbed by changing one or more values of one or more design parameters. The design parameters from the inverse model or the previously perturbed designs are perturbed. The amount and / or direction of the change can be determined based on a pattern. The change can be based on an offset and / or percentage change of the parameter.

[0060] In one embodiment, the optimization perturbs based on a previous change or perturbation. The gradient based on the objective is used. The amount of change of the objective value with respect to the previous change is used to select the amount of change and / or the parameter to be changed. For example, change one or more of the initial designs with higher efficiency. The amount and / or direction of the change are based on the comparison of the objective value and the design parameter value differences. Perturb the design to maximize the objective maximization criterion. A search pattern can be used, such as perturbing each value or subset of the design parameter values of each design sequentially.

[0061] In one example, for each initial seed, an efficiency gradient is evaluated with respect to an input value (I). A forward model (f) trained for an efficiency objective is used in the evaluation. The gradient is given by:

[0062]

[0063] The gradient is the ratio of the change in efficiency performance to the change in the design value. A gradient value greater than 0 indicates which inputs will result in an efficiency improvement, and thus those values are selected for change. The amount of perturbation can be limited, such as by a maximum 1% perturbation scaled according to the absolute gradient value of the gradient greater than 0. Other limits or amounts of change can be used.

[0064] As indicated by the arrows from action 25 to action 23, the optimization tests each perturbation with a simulation. The perturbed design is input into a machine learning simulation model, which outputs values of performance variables, including the objective (e.g., efficiency) for each perturbed design. Using the forward model, the perturbed design is evaluated according to the objective or criterion.

[0065] The optimization (e.g., perturbation and simulation) can be repeated until the gradient is below a threshold or until a repetition count threshold is reached. Repetition occurs until the gradient for each input is less than the threshold and / or the number of iterations (e.g., 100) does not exceed the threshold. A finite number of iterations (e.g., perform 500 times and then stop) can be used without other termination criteria.

[0066] Compared with genetic algorithms, using machine learning for generation and simulation increases the speed of exploring design options. The optimization evaluates the design and moves the design towards the optimal design.

[0067] During optimization, the design is fed into the forward model, and the result is tested against the objective provided by the user in action 26. The processor determines the error for each of one or more perturbed designs and / or the initial seed design. The error is between the simulated performance value and the user-specified or designated performance value. The value of the limited and constrained performance parameter is compared with the value of the same parameter in the simulation.

[0068] Any comparison can be made, such as measuring the average percentage difference. A weighted combination can be used, which has weights based on parameter importance and / or other similarity metrics. In one embodiment, the relative error is determined. The output design optimized by perturbation is run through the simulation model for evaluation. The simulated value of the performance parameter is evaluated according to the user input objective or according to the specification. The relative error is calculated and gives an estimate of the deviation of the generated design from the objective. If the error is unacceptable (e.g., above the threshold), the user can trigger more designs until the design is acceptable.

[0069] In operation 27, the processor selects one or more designs based on the errors of the various designs. The selection can be of the initial seed design and / or the perturbed design. A subset or a single design of the test or simulation design can be selected.

[0070] The selection is based on simulation. The errors determined in operation 26 use the values of the performance parameters from the forward model. The errors are used for selection, such as selecting the design with the relatively smallest error. Weighted errors can be used, such as weighting certain performance parameters (e.g., requirements or constraints) more than other parameters.

[0071] In one embodiment, the errors and the target performance parameter values are used together for selection. A set of designs and / or errors with the highest or lowest target values (e.g., highest efficiency) are used. Alternatively, designs with target values above or below a threshold are used. For these target-based selections, one or more designs with the relatively lowest error are selected. A set of numbers or errors below the threshold will be output as the selected designs. In other embodiments, the errors are used to determine a subset (e.g., select designs with errors below a threshold), and the selection is completed by selecting the low-error designs with the maximum or minimum target values. In one embodiment, one or more designs with the highest efficiency with zero error with respect to the requirements and constraints are selected.

[0072] In one embodiment, after perturbing along the optimized design direction, several candidate designs are selected (e.g., three or any number of targets with values above or below a threshold). In other embodiments, the best design is the output based on a weighted combination of the errors and the target parameter values. In other embodiments, when the target parameter is included in the parameters used for the errors, the errors are used alone.

[0073] In operation 28, the one or more selected designs are stored in a memory. The one or more designs are stored for evaluation by one or more engineers, such as running an entity-based simulation on the one or more designs. The designs can be stored for production. One or more objects (e.g., gas turbine engines) are manufactured according to the selected designs. The manufactured objects can be tested to compare with the simulated performance.

[0074] The one or more selected designs can be used as the initial designs in a further design process. For example, for high-fidelity operation, the design output of machine learning-based optimization is used for seed generation based on the initial constraints rather than a random seed to significantly accelerate the generation of optimized designs using traditional or manually controlled optimization techniques. The one or more selected designs are manually changed and / or used in an entity-based simulator for further evaluation.

[0075] Any design approved by an expert can be used to further improve machine learning. Retraining a generative model or an inverse model and / or a simulator or a forward model using the approved design provides additional training data. Alternatively, performing incremental or online machine learning to update the machine learning network based on the approved design.

[0076] Figure 5 An embodiment of a system for a machine learning-based design is shown. Two deeply trained neural networks 54, 56 are applied. The system implements Figure 2 the method, but other methods can also be implemented. In an alternative embodiment, the system is used to train the networks 54, 56. The system can implement Figure 1 the method, but another machine learning can be used.

[0077] The system includes a design processor 50 and a memory 52. Additional, different, or fewer components can be provided. For example, a network or a network connection is provided, such as for networking the design processor 50 and the memory 52. In another example, a display and / or a user interface (e.g., a display and user input devices - a mouse and a keyboard) are provided for interaction with learning and / or design.

[0078] The design processor 50 and / or the memory 52 are part of an engineering workstation. Alternatively, the design processor 50 and / or the memory 52 are part of a separate computer or server.

[0079] The memory 52 is a graphics processing memory, a video random access memory, a random access memory, a system memory, a cache memory, a hard disk drive, an optical medium, a magnetic medium, a flash drive, a buffer, a database, a combination thereof, or other currently known or later developed non-transitory storage devices for storing data. The memory 52 is part of the design processor 50, part of a computer associated with the design processor 50, part of a database, part of another system, or an independent device.

[0080] The memory 52 stores training data, specifications, parameters, parameter values, an inverse machine learning network 54 and / or a forward machine learning network 56. The memory 52 can store other information. For example, storing the values of input feature vectors, the values of output vectors, the connections / nodes / weights / convolution kernels of the deep machine learning networks 54, 56, the calculated values 54, 56 when applying the networks, selections, errors, optimization information, gradients, and / or other data from the design. The design processor 50 can use the memory 52 to Figure 2 temporarily store information during the execution of the

[0081] Memory 52 or other memory may alternatively or additionally be a non-transitory computer-readable storage medium that stores data representing instructions executable by the programming design processor 50. On the non-transitory computer-readable storage medium or memory, such as a cache memory, buffer, RAM, removable media, hard disk drive, or other computer-readable storage medium, instructions are provided for implementing the processes, methods, and / or techniques discussed herein. Non-transitory computer-readable storage media include various types of volatile and non-volatile storage media. In response to one or more sets of instructions stored in or on the computer-readable storage medium, the functions, acts, or tasks shown in the figures or described herein are performed. The functions, acts, or tasks are independent of the particular type of instruction set, storage medium, processor, or processing strategy and may be performed by software, hardware, integrated circuits, firmware, microcode, etc., operating alone or in combination. Similarly, processing strategies may include multiprocessing, multitasking, parallel processing, etc.

[0082] In one embodiment, the instructions are stored on a removable media device for reading by a local or remote system. In other embodiments, the instructions are stored at a remote location for transmission over a computer network or telephone line. In other embodiments, the instructions are stored within a given computer, CPU, GPU, or system.

[0083] The design processor 50 is an artificial intelligence processor, such as a general-purpose processor, central processor, control processor, graphics processor, digital signal processor, application-specific integrated circuit, field-programmable gate array, massively parallel processor, a processor specifically designed to implement a machine learning network processor, digital circuits, analog circuits, combinations thereof, or other currently known or later developed devices for generating and evaluating engineering designs. The design processor 50 is a single device or multiple devices operating in series, parallel, or separately. The design processor 50 may be the main processor of a computer such as a laptop, server, or desktop computer, or may be a processor for handling some tasks in a larger system. The design processor 50 is configured by instructions, firmware, design, hardware, and / or software to perform the actions discussed herein.

[0084] The design processor 50 is configured by instructions or software stored in the memory 52 to perform inverse modeling of design parameters according to design specifications. The inverse modeling uses a machine-trained neural network. Requirements, constraints, target parameters, and / or other parameters representing the desired operating characteristics of the design are used for inverse modeling of design parameters, such as the settings of design variables. Given the values of the specifications, the inverse model generates design parameters for one or more (e.g., multiple) possible designs.

[0085] The design processor 50 is configured by instructions or software stored in the memory 52 to optimize design parameters based on a simulation of the design. Using design parameters from inverse modeling, the design processor 50 simulates the performance or operation of the design using a neural network trained as a forward model on the same machine or a different machine. By varying the design, the performance criteria of the design can be maximized or minimized while still meeting or being within the tolerances of the requirements and / or constraints of the specifications. By perturbing the design parameters of multiple possible designs from the inverse model, optimization can generate additional possible designs.

[0086] The design processor 50 is configured by instructions or software stored in the memory 52 to select one of the possible designs as the design for output of the design parameters. More than one possible design can be selected for output. The selected design is then output to, for example, a memory and / or a display. The output design can be changed and / or used for the manufacture of a designed object. Due to the use of two machine learning networks in design creation, the manufactured object is more likely to provide optimal performance.

[0087] Although the invention has been described above with reference to various embodiments, it should be understood that many changes and modifications can be made without departing from the scope of the invention. Accordingly, the foregoing detailed description is to be regarded as illustrative rather than restrictive, and it should be understood that the following claims, which are intended to define the spirit and scope of the invention, include all equivalents.

Claims

1. A method for generative design of a turbine by an artificial intelligence processor, the method comprising: generating, by the artificial intelligence processor, one or more designs of the turbine to be designed, the generation being performed by a first machine learning network in response to an input of first values of design requirements, constraints, and objectives for optimizing the efficiency of the turbine to be designed; simulating, by the artificial intelligence processor, the operation of each of the one or more designs by a second machine learning network, the simulation providing second values of the design requirements, the constraints, and / or the objectives; perturbing the one or more designs; repeating the simulation for the perturbed one or more designs; determining an error of the perturbed one or more designs, the error being the error between the second values of the simulation and the first values; and selecting, based on the error, at least one of the perturbed or generated one or more designs; storing at least one of the selected perturbed or generated one or more designs for generating the turbine; and receiving input values of performance parameters, wherein the performance parameters include ambient temperature, turbine shaft power, turbine exhaust temperature, and inlet flow rate of the burner cooling section, and the constraints are the performance parameters.

2. The method according to claim 1, wherein, the generation includes generating using the first machine learning network including an inverse model.

3. The method according to claim 1, wherein, the generation includes generating using the first machine learning network including a generative adversarial network or a mixture density network.

4. The method according to claim 1, wherein, the generation includes generating in response to the objective being one of the requirements or the input of the constraint.

5. The method according to claim 1, wherein, the generation includes generating in response to the objective being the input of efficiency.

6. The method according to claim 1, wherein, the simulation includes simulating by the second machine learning network including a fully connected residual network.

7. The method according to claim 1, wherein, the simulation includes simulating by the second machine learning network including a prediction model.

8. The method according to claim 1, wherein, the perturbation includes perturbing according to a gradient based on the objective.

9. The method according to claim 1, wherein, the repetition includes repeating the perturbation and the simulation until the gradient is lower than a first threshold or until a second threshold number of repetitions occurs.

10. The method according to claim 1, further comprising selecting a subset of the perturbed one or more designs based on the simulation.

Images