Devices and methods for accelerating physics-based simulations using artificial intelligence

Abstract

A method and device for performing a physics-based simulation are provided. The processing device includes memory and a processor. The processor is configured to perform the physics-based simulation by: performing a portion of the physics-based simulation; training a neural network model based on the results of performing the first portion of the physics-based simulation; performing inference processing based on the results of training the neural network model; and providing predictions based on the inference processing as input back to the physics-based simulation.

Description

Technical Field

[0001] (Cross - Reference to Related Applications) This application claims the priority of the pending U.S. patent application Ser. No. 18 / 344,544, filed on Jun. 29, 2023, entitled "DEVICE AND METHOD FOR ACCELERATING PHYSICS - BASED SIMULATIONS USING ARTIFICIAL INTELLIGENCE", the entire disclosure of which is incorporated herein by reference.

Background Art

[0002] Physical - based simulations have been used for decades to make predictions about physical processes and make decisions based on those predictions. For example, physical - based simulations are used in numerical weather prediction and computational fluid dynamics for aircraft design. These physical - based simulations have benefited from years of improvement in computational algorithms (e.g., algorithms implemented via software) and hardware and have become a standard tool in many scientific fields, along with experiments in laboratories and in the field.

[0003] A more detailed understanding can be obtained from the following description given by way of example together with the accompanying drawings.

Brief Description of the Drawings

[0004] [Figure 1] It is a block diagram of an exemplary computing device capable of implementing one or more features of the present disclosure. [Figure 2] It is a block diagram of a device showing additional details regarding the execution of a processing task on an acceleration processing device according to an example. [Figure 3] It is a flowchart showing an exemplary method 300 for accelerating a physics - based simulation according to the features of the present disclosure. [Figure 4] This figure shows a comparison of the time used to perform a physics-based simulation without using a neural network, and the time used to perform a physics-based simulation in combination with a physics-based simulation using a neural network, according to one embodiment. [Modes for carrying out the invention]

[0005] Physical-based simulations are extremely useful for making predictions and decisions about physical processes, but they often require very large supercomputers with hundreds or thousands of computing nodes and millions of computing cores. For example, the power required to run the largest supercomputers can exceed 10 MW.

[0006] Physically based simulations can be highly accurate, but they are also expensive. For example, the grid-based or mesh-based numerical methods used in physically based simulations require numerous floating-point operations, sometimes taking weeks on the largest computers to reach the final state of the simulation. The high cost of physically based simulations often prevents their use in predicting and determining various physical processes (e.g., designing new machinery, designing new materials, or making social decisions).

[0007] Furthermore, physically based simulations can lead to inefficient use of hardware. For example, the slowdown in single-core performance improvements in recent years has forced physically based simulations to scale out to numerous cores, requiring data communication across many nodes. Often, more time is spent moving data (e.g., between memory and processor cores, or between nodes over a network) than performing the actual work (e.g., performing calculations using the data). Even at the node level, grid-based and mesh-based computations typically suffer from limited memory bandwidth (i.e., a low ratio of calculations performed per unit of data moved from memory to registers).

[0008] Artificial intelligence is the technology that enables machines (e.g., computers) to imitate or simulate human intelligence. Machine learning, a subfield of artificial intelligence, allows machines to learn from experience (e.g., training) and improve in order to solve complex problems.

[0009] In this specification, the term "machine learning network" is used interchangeably with the terms "neural network," "machine learning neural network," "deep learning neural network," and "machine learning model."

[0010] Machine learning neural networks are widely used in various techniques (e.g., image classification) to make predictions or decisions for specific tasks (e.g., whether an image contains a specific object). A neural network typically consists of multiple layers. In each layer, a filter is applied to the previous layer, and the result of each layer is known as an activation or feature map. The first and last layers in a neural network are known as the input and output layers, respectively, while the layers between the first and last layers are known as hidden layers. A machine learning model is trained to make predictions or decisions for specific tasks (e.g., whether an image contains a specific object). During training, the model is exposed to different data. In each hidden layer, the model learns essential feature maps from the data. In the output layer, the model makes predictions, receives feedback on the accuracy of those predictions, and updates itself.

[0011] Features of this disclosure include devices and methods for accelerating a physics-based simulation by using a machine learning network to perform a portion of the simulation. A portion of the physics-based simulation is replaced by a machine learning neural network model, and the system switches back from the neural network model to the physics-based simulation (for example, the neural network is trained based on the results of a portion of the physics-based simulation, the inference process is performed based on the trained neural network model, and the predictions obtained from the inference process are returned to the physics-based simulation as initial conditions).

[0012] Machine learning networks are used as efficient approximations of physics-based simulations, making predictions using fewer floating-point operations than physics-based simulations. The approximation quality of the simulation is tuned to the target accuracy of a particular simulation (e.g., a specific use case). Therefore, a particular application (use case) can be accelerated by using machine learning networks in part of a physics-based simulation while maintaining the target accuracy.

[0013] Furthermore, machine learning networks utilize computing hardware (e.g., memory and registers) more efficiently than physical-based simulations because they have a smaller footprint (i.e., smaller model size) and greater computational power than physical-based simulations. In other words, the ratio of computations performed per unit of data moved from memory to registers is higher when using machine learning networks than when using data movement in physical-based simulations.

[0014] A processing device for performing a physics-based simulation is provided, including memory and a processor. The processor is configured to perform the physics-based simulation by: performing a portion of the physics-based simulation; training a neural network model based on the results of performing the first portion of the physics-based simulation; performing inference processing based on the results of training the neural network model; and providing predictions as input back to the physics-based simulation based on the inference processing.

[0015] A method for performing a physics-based simulation is provided, which includes: performing a portion of the physics-based simulation; training a neural network model based on the results obtained from performing the first portion of the physics-based simulation; performing inference processing based on the results of training the neural network model; and providing predictions as input back to the physics-based simulation based on the inference processing.

[0016] A method for performing a physics-based simulation is provided, which includes: performing a first part of the physics-based simulation; training a neural network based on the results of the first part of the physics-based simulation to produce a trained neural network model; performing a second part of the physics-based simulation while the neural network model is being trained; performing inference processing based on the results of the trained neural network model; and, once the execution of the inference processing is complete, providing the final prediction of the inference processing to a third part of the physics-based simulation. One or more predictions about the physical process are generated from the results of the physics-based simulation.

[0017] Figure 1 is a block diagram of an exemplary computing device 100 that can implement one or more features of the present disclosure. In various examples, the computing device 100 is, but is not limited to, a computer, a gaming device, a handheld device, a set-top box, a television, a mobile phone, a tablet computer, or any other computing device. The device 100 includes, but is not limited to, one or more processors 102, memory 104, one or more auxiliary devices 106, and storage devices 108. An interconnection 112, which may be a bus, a combination of buses, and / or any other communication component, links the processor(s) 102, memory 104, auxiliary devices(s) 106, and storage devices 108 in a communicative manner.

[0018] In various alternative configurations, the processor(s) 102 includes a central processing unit (CPU), a graphics processing unit (GPU), a CPU and GPU located on the same die, or one or more processor cores, each of which may be a CPU, GPU, or neural processor. In various alternative configurations, at least a portion of the memory 104 is located on the same die as one or more of the processor(s) 102, for example, on the same chip or in an interposer configuration, and / or at least a portion of the memory 104 is located independently of the processor(s) 102. The memory 104 includes volatile or non-volatile memory (e.g., random access memory (RAM), dynamic RAM, cache).

[0019] The storage device 108 includes fixed or removable storage devices (e.g., hard disk drives, solid-state drives, optical discs, flash drives, etc.). The auxiliary device(s) 106 includes, but are not limited to, one or more auxiliary processors 114 and / or one or more input / output ("IO") devices. The auxiliary processor(s) 114 include, but are not limited to, processing units capable of executing instructions such as a central processing unit, graphics processing units, parallel processing units capable of performing computational shader operations in single-instruction multiple data formats, multimedia accelerators such as video encoding or decoding accelerators, or any other processor. Any auxiliary processor(s) 114 can be implemented as a programmable processor that executes instructions, a fixed-function processor that processes data according to fixed hardware circuits, a combination thereof, or any other type of processor. In some examples, the auxiliary processor(s) 114 include an accelerated processing device (APD) 116. In addition, although the processor(s) 102 and APD116 are shown separately in Figure 1, in some examples the processor(s) 102 and APD116 may be on the same chip.

[0020] One or more I / O devices 118 include one or more input devices such as a keyboard, keypad, touchscreen, touchpad, detector, microphone, accelerometer, gyroscope, bioscanner, or network connectivity (e.g., a wireless local area network card for transmitting and / or receiving wireless IEEE 802 signals), and / or one or more output devices such as a display, speaker, printer, haptic feedback device, one or more lights, antenna, or network connectivity (e.g., a wireless local area network card for transmitting and / or receiving wireless IEEE 802 signals).

[0021] Figure 2 is a block diagram of a device illustrating additional details regarding the execution of a processing task on the APD116, in an example. The processor 102 maintains one or more control logic modules for execution by the processor(s) 102 within the system memory 104. The control logic modules include the operating system 120, the driver 122, and the application 126, and may optionally include other modules not shown. These control logic modules control various aspects of the operation of the processor(s) 102 and the APD116. For example, the operating system 120 communicates directly with the hardware and provides an interface to the hardware for other software running on the processor(s) 102. The driver 122 controls the operation of the APD116 by providing an application programming interface (API) to software (e.g., application 126) running on the processor(s) 102, thereby accessing various functions of the APD116. The driver 122 also includes a just-in-time compiler that compiles shader code into shader programs for execution by the processing components of the APD 116 (such as the SIMD unit 138, which will be described in more detail below).

[0022] The APD116 executes commands and programs for selected functions, such as graphics and non-graphics calculations, which may be suitable for parallel processing. Based on commands received from processor(s) 102, the APD116 may be used to perform graphics pipeline calculations such as pixel calculations, geometric calculations, and rendering images to display devices (e.g., any of the I / O devices 118). The APD116 may also perform calculations that are not directly related to graphics calculations, or calculations that are not part of the "normal" information flow of the graphics processing pipeline, or calculations that are not at all related to graphics calculations (sometimes called "GPGPU" or "general-purpose graphics processing unit"), such as calculations related to video, physical simulations, computational fluid dynamics, or other tasks.

[0023] The APD116 includes a computing unit 132 (collectively referred to herein as a “programmable processing unit”) which includes one or more SIMD units 138 configured to perform operations in a parallel manner according to the SIMD paradigm. The SIMD paradigm is one in which multiple processing elements share a single program control flow unit and program counter, and thus execute the same program, but can execute that program with different data. For example, each SIMD unit 138 includes 16 lanes, and each lane executes the same instruction simultaneously with other lanes in the SIMD unit 138, but can execute that instruction with different data. Lanes can be switched off using predication if not all lanes need to execute a given instruction. Predication can also be used to execute programs with branch control flows. More specifically, conditional branch instructions or control flows enable programs with other instructions based on calculations performed by individual lanes to follow any control flow by predication of lanes corresponding to control flow paths not currently being executed and serial execution of different control flow paths.

[0024] The basic unit of execution in the computing unit 132 is a work item. Each work item represents a single instantiation of a shader program that is executed in parallel in a particular lane of a wavefront. Work items can be executed simultaneously as a "wavefront" on a single SIMD unit 138. Multiple wavefronts can be included in a "work group", which includes a collection of work items designated to execute the same program. A work group can be executed by executing each of the wavefronts that make up the work group. Wavefronts can be executed continuously on a single SIMD unit 138 or partially or fully in parallel on various SIMD units 138. A wavefront can be considered an instance that executes a shader program in parallel, but each wavefront includes multiple work items that are executed simultaneously on a single SIMD unit 138 according to the SIMD paradigm (e.g., one instruction control unit that executes a stream of the same instructions for multiple data). The command processor 137 exists within the computing unit 132 and launches wavefronts based on work (e.g., execution tasks) that are waiting for completion. The scheduler 136 is configured to perform operations related to scheduling various wavefronts on different computing units 132 and SIMD units 138.

[0025] The parallelism provided by the computing unit 132 is suitable for graphics-related operations such as pixel value calculation, vertex transformation, tessellation, geometry shading operations, and other graphics operations. Therefore, the graphics processing pipeline 134 that receives graphics processing commands from the processor(s) 102 provides computing tasks to the computing unit 132 for parallel execution.

[0026] Also, the computing unit 132 is also used to perform computational tasks that are not related to graphics or are not performed as part of the "normal" operations of the graphics processing pipeline 134 (e.g., custom operations performed to supplement the operations of the graphics processing pipeline 134). An application 126 or other software executed on the processor(s) 102 transmits a program (often referred to as a "compute shader program" and which can be compiled by the driver 122) that defines such computational tasks to the APD 116 for execution. Although the APD 116 having the graphics processing pipeline 134 is shown, the teachings of the present disclosure are also applicable to the APD 116 that does not have the graphics processing pipeline 134.

[0027] The APD 116 is configured to perform various functions to accelerate a physics-based simulation by using a machine learning network to execute a part of the physics-based simulation while maintaining the target accuracy, and to utilize the computing hardware more efficiently. For example, as described in more detail herein, the APD 116, without limitation, executes a part of the physics-based simulation, trains (or retrains) a neural network model based on the results of a part of the physics-based simulation, performs an inference process for making a prediction based on the results of the trained neural network model, and is configured to return the prediction to a part of the physics-based simulation.

[0028] Also, the APD 116 is configured to make various determinations to accelerate the physics-based simulation, such as determining whether to train or retrain a neural network, whether to retrain a neural network model based on new simulation data, or whether to use a previously trained neural network model as an initial estimate in a new training phase based on new data.

[0029] Figure 3 is a flowchart illustrating an exemplary method 300 for accelerating a physics-based simulation according to the features of the present disclosure. Each step shown in method 300 of Figure 3 is performed by an accelerating processor (e.g., APD116 shown in Figure 2).

[0030] An exemplary method 300 is illustrated using the timing diagram shown in Figure 4. The top of Figure 4 shows a baseline workflow 402 that performs a physics-based simulation without using a neural network, and the bottom of Figure 4 shows a hybrid workflow 404 that performs a physics-based simulation by combining the physics-based simulation with a neural network in accordance with the features of this disclosure. Figure 4 shows a comparison of the time used to perform a physics-based simulation without using a neural network (i.e., the baseline workflow 402 shown in the top of Figure 4) and the time used to perform a physics-based simulation according to the hybrid workflow 404.

[0031] As shown in Figure 4, both the baseline workflow 402 and the hybrid workflow 404 are executed using 1500 steps. The total number of steps shown in Figure 4 (1500) is just an example. A feature of this disclosure is that the hybrid workflow 404 may include any number of steps other than the 1500 steps shown in Figure 4. In addition, in the example shown in Figure 4, the hybrid workflow 404 includes executing different parts of a physically based simulation, each containing 500 steps, and an inference phase part containing 500 steps. However, a feature of this disclosure is that each part of the hybrid workflow 404 may include executing any number of steps other than the 500 steps shown in Figure 4.

[0032] As shown in block 302, method 300 includes performing a portion of a physics-based simulation (e.g., by APD116). The portion of the physics-based simulation is performed over several steps (e.g., by APD116). For example, as shown in Figure 4, the first portion 406 of the physics-based simulation is performed over 500 steps, starting at time t0 and ending at time t1. During the execution of any portion of the physics-based simulation (e.g., the first portion 406 of the physics-based simulation), simulation results in space and time, or a portion of the simulation results, are stored in memory. For example, subsamples in space and time may be stored in memory. In addition to or instead of this, certain state variables may be stored in memory, while other state variables are not.

[0033] As shown in decision block 304, method 300 includes determining whether there is an additional part of the physics-based simulation to be performed or additional training to be performed. Depending on whether there is no additional part of the physics-based simulation to be performed and no additional training to be performed on the neural network (decision "no"), the method terminates in block 306.

[0034] However, depending on at least one of the additional parts of the physics-based simulation that are determined to be performed and the additional training of the neural network that are determined to be performed (a "yes" decision), the method proceeds to block 308 to train (or retrain, as will be explained in more detail below) the neural network based on the results of the parts of the physics-based simulation. For example, if the accelerator (e.g., APD116) has finished performing some steps for the first part 406 of the physics-based simulation at time t1 and there are still additional parts of the physics-based simulation to be performed (i.e., the second part 410 and the third part 414), then in block 308, the neural network model 408 is trained (e.g., by APD116) based on the results of the first part 406 of the physics-based simulation.

[0035] Parts of a physics-based simulation may terminate based on, for example, one of several different criteria (and the decision is made in block 304). For example, as shown in Figure 4, the first part 406 of the physics-based simulation terminates at time t1 after completing several steps (e.g., 500 steps). Alternatively, a part of the physics-based simulation may terminate based on, for example, a predetermined amount of time elapsed since the start of that part of the physics-based simulation, but not limited to this, or based on the decision that the results of that part of the physics-based simulation are sufficient to train a neural network (and the next part begins).

[0036] As shown in Figure 4, the training of the neural network model 408 is performed (e.g., by APD116) to start at time t1 and end at time t2. Experimental data (e.g., laboratory data or field data) can also be used to train the neural network model.

[0037] As further shown in Figure 4, while the neural network model 408 is trained between time t1 and time t2, the second part 410 of the physics-based simulation is also executed between time t1 and time t2. Additional data generated by the second part 410 of the physics-based simulation may be provided to the neural network model 408 as real-time input during training. Alternatively, the additional data generated by the second part 410 of the physics-based simulation may be used to validate the trained neural network model 408, i.e., the first 500 steps generate the training dataset, and the additional data form the validation set.

[0038] As shown in block 310, method 300 includes performing an inference process (i.e., an inference phase that makes one or more predictions) based on the results of a trained neural network model. For example, as shown in Figure 4, an inference process (i.e., inference phase 412) is performed between time t2 and time t3 based on the results of training the neural network model 408. In the example shown in Figure 4, inference phase 412 is performed over the same number of steps (500 steps) as the first part 410 of the physically based simulation. Thus, as shown by the smaller time interval between time t2 and time t3 (compared to the larger time interval between time t0 and time t1), inference phase 412 (based on the results of the trained neural network) is performed much faster over the same number of steps as the physically based simulation.

[0039] Inference phase 412 continues until it reaches a number of steps (e.g., specified by the user via user input), until the prediction no longer satisfies a physical relationship (e.g., defined by user input), or until the prediction uncertainty is equal to or greater than a threshold prediction uncertainty (e.g., specified by user input).

[0040] As shown in block 312, method 300 includes returning the predictions from the inference process (e.g., the last prediction) to a physics-based simulation once the execution of the inference process is complete. For example, as shown in Figure 4, once the execution of the inference phase 412 is completed at time t3, the last prediction obtained from the inference phase 412 is returned to the third part 414 of the physics-based simulation as an initial condition.

[0041] Next, as shown in Figure 4, method 300 reverts to block 302 and the next part of the physical-based simulation (for example, the third part of the physical-based simulation in this example) from time t3 to time t4, based on the initial conditions provided to the third part 414 of the physical-based simulation obtained from the inference phase 412. In the example shown in Figure 4, the third part 414 of the physical-based simulation is also executed over 500 steps.

[0042] During the execution of the third part 414 of the physics-based simulation, the third part 414 of the physics-based simulation corrects for erroneous predictions arising from the inference phase 412.

[0043] In decision block 304, a decision is made again (for example, by APD116) regarding whether there are any additional parts of the physics-based simulation to be performed or any additional training of the neural network to be performed. For example, if, after performing the third part 414 of the physics-based simulation, there are no additional parts of the physics-based simulation and no additional training of the neural network to be performed (decision "no"), the method terminates in block 306.

[0044] Although not shown in Figure 4, new simulation data may be generated in some cases to retrain the neural network model 408 and to improve the accuracy of the predictions made by the neural network model 408. For example, after the third part 414 of the physical-based simulation is performed in block 302, data generated during the third part 414 of the physical-based simulation may be used to retrain the neural network model 408 in block 308, depending on whether additional training of the neural network model 408 is decided in the decision block 304.

[0045] Alternatively, a new neural network model can be trained from scratch (i.e., without data generated during the third part of the physics-based simulation 414). For example, if the amount of change in the simulated physical behavior of the system exceeds a change threshold and the previously trained model should not be used, it is determined (e.g., by APD116) that a new model should be retrained. Or, the previously trained neural network model is saved and used as an initial estimate in a new training phase based on new data.

[0046] After the neural network model is retrained using the data generated during the third part 414 of the physics-based simulation (or a new neural network model is trained from scratch), the inference process is performed in block 310 so that the predictions are provided to the physics-based simulation (for example, as input to the third part 414 of the physics-based simulation), and then the method returns to block 302 to re-execute the final third part of the physics-based simulation.

[0047] After the physics-based simulation is complete, in block 306, one or more predictions about the physical process (e.g., predictions for designing a new machine, a prediction for designing a new material, or a prediction for making a social decision) are determined (e.g., by APD116).

[0048] As shown in Figure 4, the physics-based simulation is performed in less time (i.e., as indicated by the time gain between time t4 and time t5) than when the physics-based simulation is performed without the training and inference process of the neural network model 408 (i.e., the baseline workflow 402), using the training and inference phase 412 of the machine learning neural network model 408 (i.e., the hybrid workflow 404).

[0049] It should be understood that many variations are possible based on the disclosures herein. Although features and elements are described above in specific combinations, each feature or element can be used alone without other features and elements, or in various combinations with or without other features and elements.

[0050] The methods provided may be implemented in a general-purpose computer, processor, or processor core. Suitable processors include, by example, general-purpose processors, dedicated processors, conventional processors, digital signal processors (DSPs), multiple microprocessors, one or more microprocessors associated with a DSP core, controllers, microcontrollers, application-specific integrated circuits (ASICs), field-programmable gate array (FPGA) circuits, any other type of integrated circuit (IC), and / or state machines. Such processors may be manufactured by configuring a manufacturing process using the results of processed hardware description language (HDL) instructions and other intermediate data (such as instructions that can be stored in a computer-readable medium), including netlists. The results of such processing may be maskwork, which is then used in a subsequent semiconductor manufacturing process to manufacture a processor that implements the features of the present disclosure.

[0051] The methods or flowcharts provided herein may be implemented in computer programs, software, or firmware incorporated into non-temporary computer-readable storage media for execution by a general-purpose computer or processor. Examples of non-temporary computer-readable storage media include read-only memory (ROM), random-access memory (RAM), registers, cache memory, semiconductor memory devices, magnetic media such as internal hard disks and removable disks, magneto-optical media, and optical media such as CD-ROM disks and digital versatile disks (DVDs).

Claims

1. A processing device, Memory and Equipped with a processor, The aforementioned processor, This involves performing a portion of a physics-based simulation, Training a neural network model based on the results of performing a portion of the aforementioned physics-based simulation, The process involves performing inference based on the training results of the aforementioned neural network model, Based on the aforementioned inference process, the prediction is provided as input back to the physics-based simulation, The system is configured to perform the aforementioned physics-based simulation. Processing device.

2. The processor is configured to provide the prediction as input to the next part of the physical-based simulation. The processing device according to claim 1.

3. The processor is configured to train the neural network model in accordance with at least one of the following: an additional portion of the physical-based simulation that is determined to be performed based on the results of performing a portion of the physical-based simulation, and an additional training of the neural network that is determined to be performed. The processing device according to claim 1.

4. In response to the decision to further train the neural network model, the processor is configured to retrain the neural network model using data generated during the portion of the physical-based simulation. The processing device according to claim 3.

5. In response to the decision to further train the neural network model, the processor is configured to train a new neural network model without using data generated during the portion of the physical-based simulation. The processing device according to claim 3.

6. The aforementioned physics-based simulation and the aforementioned inference process are performed over the same number of steps. The processing device according to claim 1.

7. The inference process allows the same number of steps to be executed in a shorter time than a portion of the physical-based simulation. The processing device according to claim 6.

8. The processor is configured to store in memory a portion of the results of executing a portion of the physical-based simulation. Some of the results mentioned above include subsamples of simulation results in space and time, and at least one of the state variables. The processing device according to claim 1.

9. A method for performing a physics-based simulation, Performing a part of the aforementioned physics-based simulation, Training a neural network model based on the results of performing a portion of the aforementioned physics-based simulation, The process involves performing inference based on the training results of the aforementioned neural network model, This includes providing the prediction, based on the inference process, as input back to the physics-based simulation, method.

10. This includes providing the aforementioned prediction as input to the following part of the physical-based simulation, The method of claim 9.

11. The process includes training the neural network model in accordance with at least one of the following: an additional portion of the physical-based simulation that is determined to be performed based on the results of performing a portion of the physical-based simulation, and an additional training of the neural network that is determined to be performed. The method of claim 9.

12. Depending on the decision to further train the neural network model, this includes retraining the neural network model using data generated during the portion of the physical-based simulation, The method according to claim 11.

13. In response to the decision to further train the aforementioned neural network model, the process includes training a new neural network model without using the data generated during the portion of the physical-based simulation, The method according to claim 11.

14. This includes training the new neural network model in response to the amount of change in the simulated physical behavior exceeding a physical behavior change threshold. The method of claim 13.

15. The aforementioned physical-based simulation and the aforementioned inference process are performed over the same number of steps. The method of claim 9.

16. The inference process allows the same number of steps to be executed in a shorter time than a portion of the physical-based simulation. The method according to claim 15.

17. The process involves storing a portion of the results obtained from performing a portion of the aforementioned physics-based simulation, wherein the portion of the results includes at least one of subsamples of the simulation results in space and time, and a portion of the state variables. The method of claim 9.

18. A method for performing a physics-based simulation, The first part of the aforementioned physics-based simulation is performed, The process involves training a neural network based on the results of the first part of the aforementioned physics-based simulation, and generating a trained neural network model. During the period in which the neural network model is being trained, the second part of the physics-based simulation is performed, This involves performing inference processing based on the results of a trained neural network model, This includes providing the prediction of the inference process to a third part of the physics-based simulation once the execution of the inference process is complete, One or more predictions regarding the physical process are generated from the results of the physical-based simulation. method.

19. This includes providing the data generated by the second part of the physics-based simulation to the neural network model while the second part of the physics-based simulation is being executed. The method of claim 18.

20. After the training of the neural network model is complete, the trained neural network model is validated using data generated by the second part of the physics-based simulation. The method of claim 18.

Images