Method, apparatus and computer program for determining a specification for an electronic device

Abstract

A method for determining a specification for an electronic device is provided. The method includes defining requirements of the electronic device. The method further includes using a trained machine learning model to predict the specification of the electronic device based on the requirements.

Description

[0001] Method, apparatus and computer program for determining a specification for an electronic device

[0002] Field

[0003] The present disclosure relates to the generation of specifications for electronic devices.

[0004] Background

[0005] In the design and development of semiconductor devices or electronic devices, particularly complex integrated circuits (ICs) such as image sensors and custom application-specific integrated circuits (ASICs), the process of defining technical specifications is a critical early step. These specifications, which outline essential design parameters (e.g., process technology, pixel count, color filter type, cell size, clock frequency), are crucial for the performance characteristics and design constraints of the chip. For image sensors, for example, specifications may include parameters such as resolution, signal-to-noise ratio, color filter array configuration, and operational modes, each contributing to the sensor’s final capability and suitability for applications ranging from consumer electronics to automotive and industrial use.

[0006] Defining technical specifications traditionally requires deep expertise and experience from semiconductor engineers who must balance various design trade-offs to meet applicationspecific requirements. This manual, experience-based process is labor-intensive and timeconsuming, as each specification must be carefully considered to avoid conflicts between design parameters that could impact performance, manufacturability, or reliability. As such, this early phase may often act as a bottleneck in the development process, delaying downstream design and prototyping activities. These delays can extend time-to-market, particularly as design iterations occur to accommodate evolving market demands or to optimize performance for a specific use case.

[0007] Another significant challenge in traditional specification processes is the limited ability to leverage historical design data effectively. While design databases exist within semiconductor organizations, accessing and utilizing this knowledge for specification development is hindered by the diversity and lack of standardization in historical data. Specifications from previous designs may vary significantly in format and structure, making it difficult to apply lessons from past designs directly. Furthermore, concerns around the confidentiality of pro- prietary data further complicate efforts to centralize specification knowledge and reuse it for new projects.

[0008] There is a demand to streamline and automate the specification development process in a way that enables efficient real-time generation, adaptation, and validation of technical specifications based on evolving user requirements and past design knowledge.

[0009] Summary

[0010] In one embodiment, a method for determining a specification for an electronic device includes defining requirements of the device and using a trained machine learning model to predict specifications based on these requirements. This method reduces manual effort and accelerates specification generation. At the same time, the probability that the specification indeed leads to a semiconductor device fulfilling the requirements may be increased. Embodiments of the method may enable real-time outputs based on the requirement.

[0011] A requirement is a desired feature or characteristic that an electronic device must have to fulfill its intended function or meet user or project needs. Requirements set the goals and constraints for the design process but do not specify how these goals will be achieved. They are typically high-level and form the foundation for subsequent technical work. Examples of requirements include a sensor that should be capable of capturing images at a resolution of 12 megapixels or higher, a device that should operate with minimal electrical noise to ensure signal integrity, or a high frame rate of at least 60 frames per second for video capture. Other examples could be the use of specific fabrication technology such as a CMOS process or the device having power consumption under 1 watt during active operation. Depending on the desired application, requirements could also be defined to meet industrial or safety standards. For example, the requirements may be defined by requiring that measurable technical parameters end up in a range required by such a standard.

[0012] A specification is a more detailed technical document or set of parameters that outline the characteristics and capabilities of an electronic device, including how it will meet the stated requirements. Specifications provide information about the device’s components, construction, and operation, enabling engineers to implement the design accordingly. For example, a specification for a sensor may state that it has a resolution of 4000 x 3000 pixels or that the device has a noise level of 3.5 dB under standard operating conditions. Specifications may also detail a device’s clock frequency, such as 500 MHz, or include the pin count, like 64 input / output pins. They could indicate the use of a 28nm CMOS process with a 4-layer metal stack or specify that the device operates at a core voltage of 1.2V and an I / O voltage of 3.3V. It goes without saying that, whenever specific dimensions or quantities are given, measurement tolerances and production tolerances are implicitly included such that mentioning a single parameter is to be understood as the disclosure of an interval of said parameter, the interval starting from the parameter reduced by a lower tolerance value and ending with the parameter increased by the upper tolerance value.

[0013] In summary, requirements can be understood to define what the device needs to achieve or support, such as high-speed data transmission, while specifications explain or define in greater detail how these requirements are met. In other words, requirements set the vision and constraints for the project, whereas specifications provide the technical blueprint to ensure those requirements are fulfilled.

[0014] In one embodiment, the method further includes predicting the performance of the electronic device based on the requirements. Predicting performance allows further early evaluation of the design’s suitability, reducing the need for multiple iterations and saving time. Performance refers to the measurable behavior or operational characteristics of an electronic device, indicating how well it functions under specified conditions. Performance metrics help evaluate the effectiveness, efficiency, and suitability of the electronic device for its intended purpose. These metrics often include aspects related to speed, power consumption, accuracy, and overall reliability. Examples of performance include how much power a device consumes during operation, such as 0.8 watts in active mode or 0.1 watts in standby mode. Another example is the response time of a sensor, which could be specified as a latency of 10 milliseconds when capturing an image. Performance might also include the data transfer rate, such as 10 Gbps for a high-speed interface, or the device’s frame rate capability, like 120 frames per second for a video capture application. Additionally, performance could be reflected in the noise level of an image sensor, such as a signal-to-noise ratio of 40 dB, or the sensor’s dynamic range, for instance, 90 dB for capturing detailed images in varying light conditions. Designing image sensors, the Point Spread Function (PSF) as an important optical characteristic may also be an important performance characteristic to predict. Overall, performance metrics help determine whether a device meets the design goals outlined in the requirements and whether it adheres to the technical specifications established for its development.

[0015] In one embodiment, the requirements defined in the method include functionalities the electronic device is expected to provide. Defining functional requirements ensures the generated specifications align closely with intended device functions. Functionalities refer to the specific tasks, operations, or capabilities that an electronic device is designed to perform. They outline what the device can do and the features it offers to fulfill its intended purpose. Functionalities are typically defined during the requirements phase and guide the development of technical specifications.

[0016] Examples of functionalities include the ability of an image sensor to capture high-resolution still images or record video at 4K resolution, optionally defining the recording capability in greater detail. For example, the capability to sample color information by chrominance subsampling (4:2:0 or 4:2:2) may be functionality at a greater level of detail. Another example is a sensor’s capability to detect low-light conditions and adjust its sensitivity automatically to improve image clarity. In communication devices, functionalities may include the ability to support wireless data transmission through Wi-Fi or Bluetooth protocols. For a microcontroller, functionalities might include processing digital signals, executing stored programs, or controlling connected hardware components through I / O pins. In addition, functionalities can also refer to power-saving modes in an electronic device, such as entering a low-power standby state when not actively in use. In more advanced electronic devices, functionalities might encompass built-in data encryption to secure information or facial recognition for user authentication. Overall, functionalities describe what an electronic device is equipped to do and form one basis for the design decisions that follow in the development process and the generation of a specification.

[0017] In one embodiment, predicting the specification involves accessing a database with information on requirements and specifications of existing devices. Accessing historical data may improve prediction accuracy by referencing past designs, enhancing precision in generated specifications. Optionally, the database further comprises information on the performance of the existing electronic devices.

[0018] In one embodiment, predicting the specification is achieved by performing a regression analysis across multiple requirements and their corresponding specifications. Regression analysis may help capture nuanced relationships among requirements, resulting in tailored and data-driven specifications.

[0019] In one embodiment, the trained machine learning model used in the method is a trained neural network. A neural network may enable the model to capture complex, non-linear relationships between requirements and specifications, improving prediction accuracy across various devices. In one embodiment, defining requirements involves selecting and modifying requirements of an existing electronic device. Modifying existing requirements simplifies the adaptation of specifications for new use cases, allowing customization without starting from scratch.

[0020] In one embodiment, the method includes providing the specifications and performance data of an existing device as inputs to the trained machine learning model. Incorporating known specifications and performance data as inputs may enhance output accuracy, providing a strong reference for similar devices.

[0021] In one embodiment, the method further includes identifying requirements that cannot be fulfilled simultaneously and generating an output signal highlighting these contradictions. Identifying contradicting requirements may prevent conflicts in specifications, reducing design iterations and guiding users toward feasible configurations.

[0022] In one embodiment, defining requirements includes using a trained language model that processes user input to generate requirements for the trained machine learning model. A language model-driven interface may simplify requirement entry, making the system accessible to users without technical expertise, and supporting a streamlined generation process.

[0023] In one embodiment, the trained language model generates an embedding of the requirements as input for the trained machine learning model. Generating embeddings for requirements facilitates precise matching with specification parameters, enhancing the relevance and accuracy of generated specifications.

[0024] In one embodiment, the trained language model proposes requirements based on user input about the intended use of the electronic device. Proposing requirements based on intended use may improve alignment with application needs, reducing the chance of omitting critical specifications for real-world use.

[0025] In one embodiment, a method for training a machine learning model to predict specifications for an electronic device includes compiling a training dataset based on existing devices, with the dataset containing both requirements met by the devices and the specifications of the devices. The machine learning model is trained using the requirements as an input and the corresponding specifications as output. Training the model on historical data improves prediction accuracy, making the system robust across a wide range of device types and capable of producing specifications that align with real-world design needs. Brief description of the Figures

[0026] Some examples of apparatuses and / or methods will be described in the following by way of example only, and with reference to the accompanying figures, in which

[0027] Fig. 1 illustrates a flow chart of an embodiment of a method for determining a specification for an electronic device;

[0028] Fig. 2 illustrates a flow chart of an embodiment of a method for generating a machine learning model to predict a specification for an electronic device;

[0029] Fig. 3 schematically illustrates a message flow within a method for determining a specification for an electronic device;

[0030] Fig. 4 schematically illustrates a message flow within a further method for determining a specification for an electronic device; and

[0031] Fig. 5 illustrates an overview of the entities involved in performing a method as illustrated by any of the figs. 1 , 3 and 4.

[0032] Detailed Description

[0033] Some examples are now described in more detail with reference to the enclosed figures. However, other possible examples are not limited to the features of these embodiments described in detail. Other examples may include modifications of the features as well as equivalents and alternatives to the features. Furthermore, the terminology used herein to describe certain examples should not be restrictive of further possible examples.

[0034] Throughout the description of the figures same or similar reference numerals refer to same or similar elements and / or features, which may be identical or implemented in a modified form while providing the same or a similar function. The thickness of lines, layers and / or areas in the figures may also be exaggerated for clarification.

[0035] When two elements A and B are combined using an “or”, this is to be understood as disclosing all possible combinations, i.e. only A, only B as well as A and B, unless expressly defined otherwise in the individual case. As an alternative wording for the same combina- tions, "at least one of A and B" or "A and / or B" may be used. This applies equivalently to combinations of more than two elements.

[0036] If a singular form, such as “a”, “an” and “the” is used and the use of only a single element is not defined as mandatory either explicitly or implicitly, further examples may also use several elements to implement the same function. If a function is described below as implemented using multiple elements, further examples may implement the same function using a single element or a single processing entity. It is further understood that the terms "include", "including", "comprise" and / or "comprising", when used, describe the presence of the specified features, integers, steps, operations, processes, elements, components and / or a group thereof, but do not exclude the presence or addition of one or more other features, integers, steps, operations, processes, elements, components and / or a group thereof.

[0037] Fig. 1 illustrates a flow chart of a method for determining a specification for an electronic device as summarized previously.

[0038] The method comprises defining requirements (100) of the electronic device.

[0039] A trained machine learning model predicts the specification (110) of the electronic device based on the defined requirements.

[0040] The method for determining a specification illustrated in fig. 1 so serves to generate the specification of the electronic device. To this end, the terms determining the specification and generating the specification can be used synonymously.

[0041] The requirements may, for example, be defined by a user from scratch. That is, the user defines all requirements without using a template. Optionally, defining the requirements may also include selecting and modifying requirements of an existing electronic device. For example, a user may indicate a device category or select an existing device. The user may then be presented with a template list of requirements for the device category or with the requirements fulfilled by select device. Said information may, for example, be presented to the user by a trained language model used to interface with a user. The user may then define the requirements given in the template list or confirm or modify the requirements of the existing device. The defined or modified requirements may, for example, also be input into the trained language model serving as a communication interface with a user. Alternatively, the requirements may be input into the machine learning model by other means. This proceeding may allow a user to adapt specifications based on prior designs, facilitating custom- ization for derivative products. The latter approach may furthermore avoid that a user misses to define a necessary requirement. For example, if the electronic device is an image sensor, the generated specifications are more likely to align with the unique parameters of the existing imaging devices. Requirements for the image sensor may include parameters such as resolution, speed, noise limits, fabrication technology, cell size, and number of wiring layers, supporting detailed customization for imaging applications. Requirements may further comprise a frame rate, a pixel size, a color filter array configuration, a spectral response, a dynamic range, a shutter type, a power consumption, a fill factor, pixel binning, and ADC bit depth. Pixel size” refers to the physical size measured in, e.g., pm. The color filter array refers to the used pattern, of the color filters on top of the pixels. The spectral response defines what range of wavelengths the sensor we can capture. The dynamic range refers to the maximum difference between bright / dark regions (e.g. in dB). The shutter type refers to the read-out mode and can, e.g., be global shutter or rolling shutter. The fill factor refers to the active area that can convert light to a signal (for a pixel) as compared to the total area. Pixel binning refers to the process of combining adjacent pixels throughout an image, by summing or averaging their values, during or after readout. According to some embodiments, predicting the specification comprises accessing a database (115) containing information on the requirements and specifications of existing electronic devices. Optionally, the database can additionally include information on the performance of existing electronic devices. This feature allows the machine learning model to leverage historical data to infer its predictions. Since the model learns how to generate the predictions, the model may also use the data of future electronic devices without the necessity to retrain the machine learning model. An alternative optional approach is performing a regression analysis (120) across multiple requirements and their corresponding specifications. A regression analysis allows to account for relationships among requirements, generating more refined and accurate specifications.

[0042] The method may also comprise generating an output of the specification (125). Optionally, contradicting requirements that cannot be fulfilled in a single electronic device may be determined (130). This may prevent specification conflicts by identifying and signaling potential issues early in the design process. In some implementations, the machine learning model is configured to identify contradicting and / or difficult input requirements, provide these to the user by means of an output signal and optionally suggest that these requirements should be adjusted. For example, contradicting (e.g. conflicting) input requirements may be two or more requirements which cannot be met simultaneously since the two or more requirements result in two items in the specification that cannot be fulfilled simultaneously within a single device. If, for example, a first requirement for an image sensor results in the specification requiring a RGB color filter pattern, while a second requirement results in the specification requiring a RYB color filter pattern, those two requirements would be contradicting requirements since a single chip of an image sensor can only have one color filter pattern. In the event of contradicting requirements, an output signal comprising the contradicting requirements may be generated to alert a user of the contradicting requirements. For example, the output signal may indicate or name all the requirements that contribute to the items of the specification that cannot be fulfilled simultaneously to leave it up to the user to change some or all the requirements until there are no more contradicting requirements determined. Additionally, or alternatively, the output signal may highlight or indicate a particular requirement that should be changed to resolve the contradiction. Additionally, or alternatively the output signal may also include a proposal as to how one or more requirements could be changed to resolve the contradiction.

[0043] Difficult input requirements that do not directly result in a contradiction may also be treated as contradicting (and hence undesired) requirements although their determination may be less straightforward. They may likewise be alerted to the user in a similar manner to modify or delete the requirements. Difficult requirements may be one or more requirements which are determined to make the design difficult, i.e. that are determined to make the prediction of associated items of the specification unreliable. An unreliable specification may be a specification that is very different to all specifications used to train the machine learning model or that are accessible by the trained machine learning model during inference. The difference between specifications may, for example, be determined using an arbitrary distance metric to determine a distance between a predicted specification and an existing specification of an existing device. Based on those considerations, difficult input requirements could, e.g., be determined by measuring the distance to existing specifications in a process where each user requirement is removed one-by-one and the distance to the nearest existing specification is monitored. When this distance becomes suddenly much smaller compared to the original distance for all requirements, then the requirement which was last removed is identified as a contradicting / difficult input requirement that the user could or should revise.

[0044] The method may also include the optional step of predicting the anticipated performance of the electronic device based on its requirements (135). Having upfront information on the performance may assist in evaluating the viability of the design before further development.

[0045] In some embodiments, defining the requirements may be performed using a trained language model (140). The language model processes user input to generate relevant re- quirements, improving the ease and precision of requirement generation. Optionally, the requirements may be represented or created as embeddings generated by the language model (145). Using embeddings of requirements as an input for the trained machine learning model may result in more context-aware specifications and furthermore provide for the possibility that the requirements can be defined by a user in different terminology. Different and eventually company specific language for identical requirements may be represented by similar embeddings so that identical or similar predictions can be generated even though the definition of the requirements were input in entirely different language.

[0046] Fig. 2 illustrates a method for training a machine learning model to predict specifications for an electronic device. The method begins by compiling a training dataset based on existing devices (200). The training dataset includes both the requirements met by these devices and the specifications they have. In other words, the training dataset comprises pairs of “[requirements, technical specifications]” and might for example be compiled for a LLM which is trained on this data, and which can be used for text-to-text conversions. Besides generating the technical specifications, the machine learning model can optionally also predict the final product performance, e.g., power consumption or chip area. In this event, the training dataset additionally comprises information on the performance, e.g. triplets of the form “[requirements, technical specifications, final product performance]”.

[0047] Once the training dataset is compiled, the method proceeds with training the machine learning model (205) using the requirements as inputs and the specifications as labelled outputs. The goal of the training is an alignment of predicted specifications with the specifications of the existing electronic device. Feedback assessing the quality of the prediction can, for example, be computed by an arbitrary error metric comparing the specifications and the predicted specifications. Optionally, the final product performance as an output of the machine learning model can be trained in the same way.

[0048] This training process allows the machine learning model to learn from prior designs, enabling it to generate accurate and relevant specifications for future electronic devices based on user-defined requirements.

[0049] A machine learning model in the context of the present disclosure can be a computational system trained to recognize patterns, make decisions, or generate predictions based on data. The model may use algorithms and mathematical structures to process and analyze input data, learning from it in order to improve the accuracy over time. In the process of generating specifications for electronic devices, a machine learning model can analyze his- torical specifications, performance metrics, or user requirements to predict optimal configurations, identify ideal parameters, or generate specifications that align with specific design goals.

[0050] Examples of machine learning models include neural networks, decision trees, and random forests. For tasks like predicting optimal design configurations based on large datasets of previous specifications, deep neural networks, which have multiple layers, may be effective. Decision trees make predictions by following a series of branching decisions based on the input data’s features. The trees can classify device parameters or select specifications that meet specific requirements. A random forest is a collection of decision trees working together, aggregating insights from various data points which may lead to more reliable and accurate predictions.

[0051] Other models, like support vector machines (SVMs) and regression models, may also be used. SVMs classify data by finding an optimal dividing line in multi-dimensional space and can be used to categorize or select design parameters that align with high-level requirements. Linear regression, which predicts continuous values, and logistic regression, which predicts categories, may also be used to estimate specific performance metrics or identify specifications that meet certain requirements.

[0052] Natural language processing (NLP) models, such as transformers or large language models, process and understand human language, making them valuable in interpreting user requirements or design intents. In this context, an NLP model might convert user-described requirements into structured data (embeddings) that can be further processed by other models to produce specifications / specific design features or parameters.

[0053] Clustering models, like K-Means, group similar data points together, which can be useful for identifying patterns in past specifications or configurations. By clustering successful designs, these models can help suggest new specifications that align with similar, proven configurations.

[0054] Each type of machine learning model can be used alone or in combination, depending on the complexity of the task and the data involved in the specification process.

[0055] In principle, specifications of arbitrary electronic devices that operate through the manipulation, processing, or transmission of electrical signals to perform specific tasks or functions can be predicted by the embodiments described. Such devices may integrate components like sensors, processors, and circuitry to interpret, process, and respond to inputs, whether from a user, another device, or the environment. Electronic devices are essential in a wide range of applications, from consumer electronics to industrial machinery, medical equipment, and telecommunications systems. The generation of specifications of electronic devices within a vast array of products and across different industries can be expedited by the embodiments described herein.

[0056] Examples of such electronic devices include smartphones, which combine functionalities like communication, internet browsing, photography, and multimedia playback in a single compact unit. Laptops and tablets serve as portable computing devices, enabling users to perform a wide array of tasks, including document processing, video conferencing, and software development.

[0057] In the realm of sensors, electronic devices such as image sensors capture and process visual information for use in cameras, smartphones, and surveillance systems, while temperature sensors monitor and regulate thermal conditions in applications ranging from Heating, Ventilation, and Air Conditioning systems to automotive engines.

[0058] Wearable devices like fitness trackers and smartwatches collect health-related data such as heart rate, steps taken, and sleep patterns, providing users with insights into their physical activity and well-being. Medical devices such as digital thermometers, glucose monitors, and electrocardiogram (ECG) machines collect and analyze physiological data for health monitoring and diagnosis.

[0059] Industrial and automotive applications rely on a range of specialized electronic devices. For example, programmable logic controllers (PLCs) automate machinery in manufacturing environments, while engine control units (ECUs) in vehicles manage engine performance and fuel efficiency. In aerospace, avionics systems rely on complex electronic devices to monitor, control, and communicate data critical to safe flight operations.

[0060] In telecommunications, electronic devices like routers, modems, and signal repeaters facilitate data transmission across networks, enabling internet connectivity and data exchange. Similarly, network servers and data storage systems in data centers process, store, and retrieve vast amounts of information for cloud computing services. Household appliances like washing machines, microwaves, and refrigerators also contain sophisticated electronics for managing cycles, temperatures, and power consumption, offering users convenience and efficiency in everyday tasks.

[0061] Fig. 3 illustrates a message flow during a design session of a user 310 making use of a method for determining a specification for an electronic device. The user 310 performs multiple iterations of the method for determining a specification for an electronic device to ultimately arrive at a specification that fulfills all his requirements. In the illustrated example, the electronic device targeted is an image sensor.

[0062] Once the user defined requirements 320 for an electronic device, the trained machine learning model 330 predicts the specifications 340 of the electronic device. Optionally, the trained machine learning model 330 also predicts the performance 350 of the electronic device. As already indicated previously, the trained machine learning model 330 may optionally access a database 360 of requirements and specifications of existing electronic devices. The database may optionally also include information on the performance of the electronic devices.

[0063] Based on the predictions, the user 310 decides as to whether he wishes to modify the requirements or to finish the process of generation of the specifications for the electronic device. For example, if the predictions of the performance 350 include a predicted energy consumption that is too high, the user may decide to redefine all or some of the requirements with the goal of reducing the predicted energy consumption.

[0064] As an alternative to the definition of the requirements from scratch or unguided, the user 310 may also decide to design an electronic device based on an existing electronic device and to modify the requirements of the existing electronic device. In this event, the specification 370 and optionally also the performance of the existing electronic device are additionally used as an input to the trained machine learning model 330.

[0065] Like fig. 3, fig. 4 schematically illustrates a message flow during a design session of a user 410 making use of a method for determining a specification for an electronic device. While the iterative process corresponds to the process illustrated in fig. 3, the requirements 420 are generated additionally using a trained language model 415. The trained language model 415 is used to query input of the user 410 ( optionally using a Chatbot interface) and to generate the requirements 420 as an input for the trained machine learning model 430. The trained language model 415 may, for example, generate the requirements as plain text. In a further embodiment, an embedding of the requirements may alternatively be used as an input for the trained machine learning model 430.

[0066] Fig. 5 schematically illustrates an overview of the functionalities involved in performing a method as illustrated by any of the figs. 1 , 3 and 4 by means of an apparatus 500 for determining a specification 540 for an electronic device.

[0067] The apparatus 500 for determining a specification for an electronic device 540 comprises an input interface 505. The input interface 505 serves to receive input from a user 512 of the apparatus 500. The apparatus 500 further comprises circuitry 510 implementing a trained machine learning model 530 configured to predict the specification 540 of the electronic device based on the requirements.

[0068] According to some embodiments, user 512 directly defines requirements 520 for an electronic device and the input interface 505 is configured to receive requirements of the electronic device to be input into the trained machine learning model 530.

[0069] According to further embodiments, circuitry 510 also implements a trained language model 515. The user 512 may interact with the trained language model 515 providing input to the trained language model 515 via the input interface. The interaction with the user 512 may be in the form of a dialogue, Hence, the trained language model 515 may also generate an output signal to query a user input. Said output signal can be made available via an output interface of the apparatus 500. The trained language model 515 may then generate the requirements 520 as an input for the trained machine learning model 530 based on the received user input.

[0070] For example, the trained language model 515 may generate an output signal querying a user 512 about an intended use of the electronic device to directly generate the requirements 520 based on the received intended use or to propose requirements to user 512.

[0071] The functionalities of the apparatus 500 may be executed by a dedicated hardware or within a general-purpose computer or by a combination of both. Likewise, the different functionalities of the apparatus 500 may be provided by a single entity or the functionalities may be distributed amongst different entities. Such an entity may comprise hardware, software, or a combination of both. Embodiments for determining a specification for an arbitrary electronic device are subsequently summarized again. Some of the aspects described in the following directed to an image sensor for the sole reason of giving a practical example, without the intent of limiting the following considerations to this particular use case. . Like in the design process of most electronic devices, a technical specification document needs to be generated early in the process for (image) sensors or for a new (semiconductor) chip. The specification contains the desired values or ranges of acceptable values for some technical aspects of the chip. For an image sensor, the specification may include the wafer process, the number of pixels, the focal length, the color filter pattern (e.g. RGB, RYB, YUV or the like), the unit cell size, the chip size, the driving clock frequency, .... Based on the technical specification, the development process is carried out and finally results in a chip. From a resultant chip, one can obtain performance metrics or a performance of the chip. The performance may comprise exact values of parameters which were only specified as ranges in the technical specification, but it may also comprise other metrics like power consumption, clock rate, chip area, number of processor cores of an optional onboard on-chip signal processing circuit etc.

[0072] Conventionally defining specifications for a new sensor / chip requires a lot of experience and expertise. It is, therefore, a time-consuming process which often can only be carried out by a limited number of engineers. As it is at the beginning of the chip design process, all later steps must wait until the technical specification is finished which might result in large timedelays and long time-to-market periods.

[0073] Methods described herein, however, can generate the specification based on previous chip designs in an interactive and real-time manner using a trained machine learning model, apparatus performing an embodiment of the method will subsequently also be named machine learning model.

[0074] The user chooses some requirements that he has: specific product features (e.g., minimum signal-to-noise ratio in low light conditions, specific imaging modes like HDR image capture) and / or specific chip features (e.g., fabrication process, resolution, color filter pattern, ...). Based on this input, the machine learning model produces a complete specification document which gives the design parameters. Alternatively or additionally, the user answers questions in a dialogue with a LLM (e.g., “Images of which objects will be taken?”, “Will the final product be used indoors or outdoors?”, “What is the intended user group?”) and then the LLM generates the input requirements for the machine learning model from the answers that the user has given. For example, the questions may prompt the user to indicate an intended use of the electronic device (e.g. “Images of which objects will be taken?”, “Will the final product be used indoors or outdoors?”, “What is the intended user group?” or the like), and the LLM generates associated input requirements based on the intended use. The dialog may be a set of questions which together build up an LLM output that meets a threshold requirement for a sufficiently detailed input requirements, where the LLM may be configured to calculate whether the threshold is met.

[0075] The machine learning model is trained on existing designs, e.g., pairs “[requirements, technical specifications]” and might for example be realized by a LLM which is trained on this data and which can be used for text-to-text conversions. Besides generating the technical specifications, the machine learning model can optionally also predict the final product performance, e.g., power consumption or chip area. For this, the machine learning model may, for example, be trained on triplets of the form “[requirements, technical specifications, final product performance]”.

[0076] A dataset with previous designs composed into a training dataset 200 is used during training of the machine learning model (Al system). A dataset with previous designs in databases 115, 360, 460 and 560 can be used during inference to find and use specifications of existing designs that are based on similar requirements than the one presently requested. It may be beneficial to compose a training dataset 200 only from publicly available data from multiple companies to train the machine learning model for subsequent use by multiple users. Likewise, it may be beneficial to provide an interface to databases 115, 360, 460 and 560 with previous designs to be used during inference such that every company or user of the machine learning model can use confidential data of own preceding designs for the prediction of new designs. Using such databases 115, 360, 460 and 560 only during inference and storing them externally, therefore, avoids the necessity to use confidential data during the training of the machine learning model. This might be preferred by semiconductor design companies that do not want to share their previous designs. The dataset in databases 115, 360, 460 and 560 may for example either be used as a classical database or as a vector database (a database of embeddings).

[0077] A vector database may be created offline and allow to quickly index the previous designs as each of them is represented by an embedding vector. Once the user inputs his requirements into the machine learning model, it is converted into an embedding used to lookup relevant previous designs / documents. This may be performed using cosine similarity between the embedding of the new query and the embedding of a previous design. The most relevant previous technical specifications may then be chosen for the user requirements and are, hence, available for the machine learning model to generate the technical specification.

[0078] A challenge in using previous design specifications is that the data is not standardized as it was manually entered. A way to deal with this problem is to use another LLM to standardize the text defining the requirements. Using in-context learning, one can describe the task of value normalization and then use the LLM to normalize the text. Such a feature normalization could be integrated as an input into an interactive chat agent or a language model used to generate requirements. The agent may generate more “friendly” feedback from feature normalized data. In addition to normalizing values, this second LLM can translate user input, such as numbers or units in a language the user knows (e.g., Japanese), into a format the model understands, such as numbers or units in English.

[0079] Based on these considerations, a specification can be generated from scratch. All necessary values in the specification are predicted based on some text prompts / requirements entered by the user. Alternatively, one can modify an existing specification to add functionality. In this case, one can alter an existing specification (called “base specification”) to add a new feature. For example, one might change an image sensor to add a new drive mode as, e.g., a low-power mode or a HDR mode or we might want to alter an image sensor and make it available for a new market (e.g., automotive) which poses other constraints (e.g., on operating temperature). Implementing this can be done using embedding space arithmetic: From the training data, one can infer a delta vector which occurs between chips that have the specific feature and chips that do not have this feature. This vector is then added to the embedding vector of the base specification and used to generate the new technical specification or to predict the new performance metrics.

[0080] In another approach, the machine learning model may be trained to find the best existing specification from a database. For example, requirements input by the user may be used to search the database for a specification which is determined to match the combination of requirements the most accurately (i.e., the closest match). The closest match may be defined differently across various embodiments and therefore determined using different methods. For example, the search maybe implemented by explicitly defining a distance function between a user query and the existing specifications, or a LLM may be used which, by means of examples, has learned to link user requirements (possibly with a certain priority which defines their weight in the search) to find the best specifications. As described earlier, in some implementations the machine learning model is configured to identify contradicting and / or difficult input requirements, provide these to the user and optionally suggest that these requirements should be adjusted. For example, contradicting (e.g. conflicting) input requirements may be two or more requirements which cannot be met simultaneously. Difficult input requirements may be one or more requirements which are determined to make the design difficult. Such difficult input requirements could, e.g., be determined by measuring the distance to existing specifications in a process where each user requirement is removed one-by-one and the distance to the nearest existing specification is monitored. When this distance becomes suddenly much smaller compared to the original distance for all requirements, then the requirement which was last removed is identified as a contradicting / difficult input requirement that the user could revise.

[0081] Similarly, an LLM may analyze the output of the machine learning model and, if the machine learning model produces two or more likely specifications, may be configured to then generate an output asking for an additional input to resolve the ambiguity. The apparatus for determining a specification for an electronic device may then output a request for the additional input guided by the LLM. To find out if there are two similar specifications (designs) for the current requirements, the LLM may, for example, slightly change the requirements and see whether the output of the machine learning model fed with the changed requirements as an input is stable or if it changes significantly. If it changes significantly, then there are several possible technical specifications. If this is the case, the LLM can be configured to generate a request for additional input. Said request can, for example, comprise questions suggesting a change of a particular requirement, based on the knowledge which of the requirements was changed and as a result caused a significantly different specification. The significance of the change may, for example, be judged using an arbitrary distance metric between arbitrary pairs of the specifications generated for each of the changed requirements.

[0082] During its use, the machine learning model may provide a proposed specification output and, with each new requirement entered by the user or each specification that he fixes, adjust its output and adapt it more to the user requirements in an iterative process.

[0083] In the following, some examples of the proposed concept are presented:

[0084] (1) A method for determining a specification for an electronic device, comprising defining requirements of the electronic device, and using a trained machine learning model to predict the specification of the electronic device based on the requirements. (2) The method according to (1), further comprising predicting a performance of the electronic device based on the requirements.

[0085] (3) The method according to (1) or (2), further comprising that the requirements define functionalities to be provided by the electronic device.

[0086] (4) The method according to any one of (1) to (3), further comprising that predicting the specification comprises accessing a database comprising information on the requirements and specifications of existing electronic devices.

[0087] (5) The method according to (4), wherein the database further comprises information on the performance of the existing electronic devices.

[0088] (6) The method according to (4) or (5), further comprising that predicting the specification comprises a regression of multiple requirements and their corresponding specifications.

[0089] (7) The method according to any one of (1) to (6), further comprising that the trained machine learning model comprises a trained neural network.

[0090] (8) The method according to any one of (1) to (7), further comprising that defining the requirements comprises selecting and modifying requirements of an existing electronic device.

[0091] (9) The method according to (7), further comprising providing the specification and the performance of the existing electronic device as an input to the trained machine learning model.

[0092] (10) The method according to any one of (1) to (9), further comprising determining contradicting requirements that cannot be fulfilled in a single electronic device, and generating an output signal comprising the contradicting requirements.

[0093] (11) The method according to (10), further comprising that defining requirements comprises using a trained language model to query a user input and to generate the requirements as an input for the trained machine learning model based on the user input. (12) The method according to (11), further comprising that the trained language model generates an embedding of the requirements as the input to the machine learning model.

[0094] (13) The method according to (11) or (12), further comprising that the trained language model generates proposed requirements based on a user input indicating an intended use for the electronic device.

[0095] (14) The method according to any one of (1) to (13), further comprising that the electronic device is an image sensor.

[0096] (15) The method according to (14), further comprising that the requirements comprise at least one of a resolution, a speed, a noise limit, a fabrication technology, a cell size, a number of wiring layers, a frame rate, a pixel size, a color filter array configuration, a spectral response, a dynamic range, a shutter type, a power consumption, a fill factor, pixel binning, and ADC bit depth.

[0097] (16) A method for generating a machine learning model to predict a specification for an electronic device, comprising compiling a training data set based on existing electronic devices, the training data set comprising requirements fulfilled by the electronic devices and the specification of the electronic devices, and training the machine learning model using the requirements as an input and the specification as a labelled output.

[0098] (17) An apparatus for determining a specification for an electronic device, comprising an input interface configured to receive requirements of the electronic device, and circuitry implementing a trained machine learning model configured to predict the specification of the electronic device based on the requirements.

[0099] (18) The apparatus according to (17), further comprising that the circuitry implementing the trained machine learning model is further configured to predict performance of the electronic device.

[0100] (19) The apparatus according to (17) or (18), further comprising that the requirements define functionalities to be provided by the electronic device.

[0101] (20) The apparatus according to any one of (17) to (19), further comprising that the circuitry implementing the trained machine learning model is further configured to access a database comprising information on requirements and specifications of existing electronic devices. (21) The apparatus according to (20), wherein the database further comprises information on the performance of the existing electronic devices.

[0102] (22) The apparatus according to (20) or (21), further comprising that the circuitry implementing a trained machine learning model performs a regression of multiple requirements and their corresponding specifications.

[0103] (23) The apparatus according to any one of (17) to (22), wherein the trained machine learning model comprises a trained neural network.

[0104] (24) The apparatus according to any one of (17) to (23), further comprising the input interface being configured to receive a selection of an existing electronic device; an output interface being configured to output the requirements of the existing electronic device; and the input interface being further configured to receive modified requirements based on the requirements of the existing electronic device as an input to the trained machine learning model.

[0105] (25) The apparatus according to (24), further comprising using the specification and the performance of the existing electronic device as an input for the trained machine learning model.

[0106] (26) The apparatus according any one of (17) to (25), further comprising that the circuitry implementing a trained machine learning model determines contradicting requirements that cannot be fulfilled in a single electronic device and generates an output signal comprising the contradicting requirements.

[0107] (27) The apparatus according any one of (17) to (26), further comprising the circuitry implementing a trained language model configured to generate an output signal querying a user input, to receive the user input, and to generate the requirements as an input for the trained machine learning model based on the received user input.

[0108] (28) The apparatus according to (27), wherein the trained language model is configured to generate an embedding of the requirements as the input to the machine learning model. (29) The apparatus according to (27) or (28), further comprising the trained language model generating an output signal querying an intended use for the electronic device; and to receive the intended use as the user input.

[0109] (30) The apparatus according any one of (17) to (29), wherein the electronic device is an image sensor.

[0110] (31) The apparatus according any one of (17) to (30), wherein the requirements comprise at least one of a resolution, a speed, a noise limit, a fabrication technology, a cell size, a number of wiring layers, a frame rate, a pixel size, a color filter array configuration, a spectral response, a dynamic range, a shutter type, a power consumption, a fill factor, pixel binning, and ADC bit depth.

[0111] (32) A computer program having program code for performing, when executed by a processor, a method comprising defining requirements of the electronic device, and using a trained machine learning model to predict the specification of the electronic device based on the requirements.

[0112] (33) A computer readable storage medium having stored thereon program code for performing, when executed by a processor, a method comprising defining requirements of the electronic device, and using a trained machine learning model to predict the specification of the electronic device based on the requirements.

[0113] The aspects and features described in relation to a particular one of the previous examples may also be combined with one or more of the further examples to replace an identical or similar feature of that further example or to additionally introduce the features into the further example.

[0114] Examples may further be or relate to a (computer) program including a program code to execute one or more of the above methods when the program is executed on a computer, processor or other programmable hardware component. Thus, steps, operations or processes of different ones of the methods described above may also be executed by programmed computers, processors or other programmable hardware components. Examples may also cover program storage devices, such as digital data storage media, which are machine-, processor- or computer-readable and encode and / or contain machineexecutable, processor-executable or computer-executable programs and instructions. Pro- gram storage devices may include or be digital storage devices, magnetic storage media such as magnetic disks and magnetic tapes, hard disk drives, or optically readable digital data storage media, for example. Other examples may also include computers, processors, control units, (field) programmable logic arrays ((F)PLAs), (field) programmable gate arrays ((F)PGAs), graphics processor units (GPU), application-specific integrated circuits (ASICs), integrated circuits (ICs) or system-on-a-chip (SoCs) systems programmed to execute the steps of the methods described above.

[0115] It is further understood that the disclosure of several steps, processes, operations or functions disclosed in the description or claims shall not be construed to imply that these operations are necessarily dependent on the order described, unless explicitly stated in the individual case or necessary for technical reasons. Therefore, the previous description does not limit the execution of several steps or functions to a certain order. Furthermore, in further examples, a single step, function, process or operation may include and / or be broken up into several sub-steps, -functions, -processes or -operations.

[0116] If some aspects have been described in relation to a device or system, these aspects should also be understood as a description of the corresponding method. For example, a block, device or functional aspect of the device or system may correspond to a feature, such as a method step, of the corresponding method. Accordingly, aspects described in relation to a method shall also be understood as a description of a corresponding block, a corresponding element, a property or a functional feature of a corresponding device or a corresponding system.

[0117] The following claims are hereby incorporated in the detailed description, wherein each claim may stand on its own as a separate example. It should also be noted that although in the claims a dependent claim refers to a particular combination with one or more other claims, other examples may also include a combination of the dependent claim with the subject matter of any other dependent or independent claim. Such combinations are hereby explicitly proposed, unless it is stated in the individual case that a particular combination is not intended. Furthermore, features of a claim should also be included for any other independent claim, even if that claim is not directly defined as dependent on that other independent claim.

Claims

ClaimsWhat is claimed is:

1. A method for determining a specification for an electronic device, comprising: defining requirements of the electronic device; and using a trained machine learning model to predict the specification of the electronic device based on the requirements.

2. The method of claim 1, further comprising: predicting a performance of the electronic device based on the requirements.

3. The method of claim 1, wherein the requirements define functionalities to be provided by the electronic device.

4. The method of claim 1, wherein predicting the specification comprises accessing a database comprising information on requirements and specifications of existing electronic devices.

5. The method of claim 4, wherein the database further comprises information on the performance of the existing electronic devices.

6. The method of claim 5, wherein predicting the specification comprises a regression of multiple requirements and their corresponding specifications.

7. The method of claim 1 , wherein the trained machine learning model comprises a trained neural network.

8. The method of claim 1 , wherein defining the requirements comprises: selecting and modifying requirements of an existing electronic device.

9. The method of claim 8, further comprising: providing the specification and the performance of the existing electronic device as an input to the trained machine learning model.

10. The method of any one of claim 1 , further comprising:determining contradicting requirements that cannot be fulfilled in a single electronic device; and generating an output signal comprising the contradicting requirements.

11. The method of claim 1 , wherein defining requirements comprises: using a trained language model to query a user input and to generate the requirements as an input for the trained machine learning model based on the user input.

12. The method of claim 11, wherein the trained language model generates an embedding of the requirements as the input to the machine learning model.

13. The method of claim 11, wherein the trained language model generates proposed requirements based on a user input indicating an intended use for the electronic device.

14. The method of claim 1 , wherein the electronic device is an image sensor.

15. The method of claim 14, wherein the requirements comprise at least one of a resolution, a speed, a noise limit, a fabrication technology, a cell size, a number of wiring layers, a frame rate, a pixel size, a color filter array configuration, a spectral response, a dynamic range, a shutter type, a power consumption, a fill factor, pixel binning, and ADC bit depth.

16. A method for generating a machine learning model to predict a specification for an electronic device, comprising: compiling a training data set based on existing electronic devices, the training data set comprising requirements fulfilled by the electronic devices and the specification of the electronic devices; and training the machine learning model using the requirements as an input and the specification as a labelled output.

17. An apparatus for determining a specification for an electronic device, comprising: an input interface configured to receive requirements of the electronic device; and circuitry implementing a trained machine learning model configured to predict the specification of the electronic device based on the requirements.

18. The apparatus according to claim 17, wherein the circuitry implementing the trained machine learning model is further configured to predict performance of the electronic device.

19. The apparatus according to claim 17, wherein the requirements define functionalities to be provided by the electronic device.

20. The apparatus according to claim 17, wherein the circuitry implementing the trained machine learning model is further configured to access a database comprising information on requirements and specifications of existing electronic devices.

Images