Neural Concept https://neuralconcept.com 3D Deep Learning Software for Enhanced Engineering Thu, 01 Apr 2021 11:00:20 +0000 en-US hourly 1 https://wordpress.org/?v=5.3.6 On the quest to break the world record: SP80 and Neural Concept are collaborating to optimize ventilated hydrofoils profiles https://neuralconcept.com/customer-cases/on-the-quest-to-break-the-world-record-sp80-and-neural-concept-are-collaborating-to-optimize-ventilated-hydrofoils-profiles/ Tue, 16 Feb 2021 11:10:07 +0000 https://neuralconcept.com/?p=4226 SP80 testing on Lake Geneva (target SPeed: 80 knots)   Neural Concept is on a mission to guide engineers towards radically better designs while speeding up their workflow. In this article, we show how SP80 managed to improve their ventilating hydrofoil performances by more than 20%, using Neural Concept Shape (NCS) in their standard design workflow.   SP80 […]

The post On the quest to break the world record: SP80 and Neural Concept are collaborating to optimize ventilated hydrofoils profiles appeared first on Neural Concept.

]]>
SP80 testing on Lake Geneva (target SPeed: 80 knots)

 

Neural Concept is on a mission to guide engineers towards radically better designs while speeding up their workflow. In this article, we show how SP80 managed to improve their ventilating hydrofoil performances by more than 20%, using Neural Concept Shape (NCS) in their standard design workflow.

 

SP80 is a team of a team of EPFL engineers and students gathered to build a boat capable of reaching 80 knots (150 kmph), propelled solely by the wind and thus break the world sailing speed record in 2022. In order to achieve this goal, the usual sailing codes have to be rethought, especially from a hydrodynamic point of view. Indeed classical sailboats foils are stuck at speeds of 50kts (around 100 kmph) due to a phenomenon called cavitation. At such speeds it becomes impossible to avoid liquid water from transforming into vapor, thus generating big instabilities that keep boats from accelerating further.

 

SP80 is therefore working on a technology called ventilation. Instead of classical tear-drop shaped foils, their foil profile will be wedged shaped. This kind of shape helps create an air bubble around the low-pressure side of the foil (the one that would cavitate on a classical foil) that is big enough to avoid any instabilities and therefore allowing the boat to break the 50kts barrier.

 

 

In order to get the best out of these ventilating hydrofoils, Charles de Sarnez, head of hydrodynamics department at SP80, has set up an optimisation framework based on cavitation tunnel tests coupled with numerical simulations, all powered by artificial intelligence thanks to Neural Concept’s software.

 

The first step was to build a numerical model capable of accurately predicting the performances of the ventilating hydrofoils. This could be done thanks to ANSYS Fluent software on the advice of CADFEM and the use of a cavitation tunnel to experimentally validate the model. The optimisation process could then start : a dataset of a several hundreds of high fidelity simulations was generated and the obtained results could then be used to train a deep-learning model, SP80 having chosen to use artificial intelligence to optimize these profiles.

 

 

Neural Concept’s deep learning model is at the heart of SP80’s hydrofoil development as it allows the team to generate and characterize hundreds of types of profiles numerically before isolating the most promising ones to have them tested experimentally in a cavitation tunnel for final validation. Thanks to Neural Concept’s model, the team can thus explore paths they could not have been able to explore in such a short period of time and make sure the best profile will be chosen. This approach has already generated a gain of more than 20% on the hydrofoil’s performances directly translating into a maximum speed increase of several kph. But ventilation remains a pretty unexplored field of hydrodynamics and a lot still has to be invented.

 

Rendering obtained thanks to Thomas Ramseier’s work during his Master project with Prof. Marco Picasso

 

Typical pressure distribution on a superventilating hydrofoil, the flat corresponds to the pressure on the ventilated side of the hydrofoil. Comparison between the Ansys Fluent simulation and NCS prediction.

 

 

 

© Aurore Kerr and Charles de Sarnez at SP80, Thomas von Tschammer at Neural Concept

The post On the quest to break the world record: SP80 and Neural Concept are collaborating to optimize ventilated hydrofoils profiles appeared first on Neural Concept.

]]>
The importance of uncertainty quantification for Deep Learning models in CAE https://neuralconcept.com/research/the-importance-of-uncertainty-quantification-for-deep-learning-models-in-cae/ Mon, 25 Jan 2021 09:25:33 +0000 https://neuralconcept.com/?p=3997 When issuing a prediction, one can question human or artificial intelligence on how much she/it is sure about the prediction. Without going into detail on the dramatic consequences of overconfident artificial agents in such situations as Autonomous Driving, this article will focus on the Neural Concept’s core competence: building predictive models that are surrogates of […]

The post The importance of uncertainty quantification for Deep Learning models in CAE appeared first on Neural Concept.

]]>
When issuing a prediction, one can question human or artificial intelligence on how much she/it is sure about the prediction. Without going into detail on the dramatic consequences of overconfident artificial agents in such situations as Autonomous Driving, this article will focus on the Neural Concept’s core competence: building predictive models that are surrogates of more resource-intensive CAE models.

 

CAE simulations are physics-based and solve numerical equations on computational domains, mostly of industrial interest (CAD shapes). We will assume that CAE yields a reasonable estimation of a real life behaviour and can therefore be assumed as “Ground Truth” (otherwise, it can be supplemented by experimental data). Similarly to the human brain, Deep Learning data-driven models learn by seeing examples and extracting information from these results. In our case, learnt examples are from CAD or various inputs (materials, boundary conditions etc.) and learnt results are CAE (or experimental). Hence, when facing a new unknown example, the model will predict the result in real-time, based on the data used for its training.

 

In general, when issuing a prediction, one can question human or artificial intelligence on how much she/it is sure about the prediction. The same is true for our Neural Network models.

 

The Deep Learning implementation of Neural Concept Shape is a tool to build and support surrogate models that produce outputs. The surrogate model (trained neural network) undergoes extensive training and testing phases, having measurables such as I1 (the mean average error) and R2 (the coefficient of determination). However, an R2 is a sort of “backwards-looking” measurable on the training/testing phase. During the neural network deployment phase, engineers would like to have a “dynamic” way to assess if the prediction is reliable. Suppose the neural network issues a signal of uncertainty above a certain predefined level. In that case, the human or artificial agent could decide to activate remediations such as submitting the neural network to new training (for efficiency, starting from the previous neural network and using Transfer Learning technology).

 

Therefore, Neural Concept has also worked extensively on uncertainty estimation, to help engineers during deployment phase facing epistemic uncertainty (degree of variation due to lack of knowledge on the model we are trying to predict). As we can see from the example given with figure 1, after uploading a given CAD geometry (here called “geometry_0001.stl”), the engineer receives from the neural network real-time predictions on the values of interest together with a confidence index from the model.

 

Another typical application is generative design. In this case, Neural Concept Shape is an agent that may create geometric shapes far out of the initial design envelope. We wish to know if the predictions  associated with the new generated shapes are still reliable, or more input/output samples that one needs for neural network re-training.

 

NCS_application

Figure 1: Application example of Neural Concept Shape for non-AI specialists (NCS /Lite), with a trained neural network embedded in a vertical application. The user can upload geometries directly from its CAD software and instantly predict the values of interest for the analysis and the degree of confidence (in this case 97%).

 

Figure 2: Using the uncertainty feature from Neural Concept Shape, we can sort the test samples by the confidence index given by the model, from lowest to highest. We show then the cumulative coefficient of determination R2 based on the sorted samples.

 

 

Masksembles: A New Methodology to Compute Uncertainty in Prediction

 

Neural Concept’s staff collaborates in research topics on top of available software capabilities, mainly with EPFL (Lausanne – Switzerland). This final section will report some activity carried out on a novel promising methodology called Masksembles.

 

Masksembles contains in itself the explanation of its technology, as follows more detailed:

 

An “ensemble”, in general, is a set of virtual copies of anything (in physics, a vessel under pressure; in AI, a neural network; in society, a collection of individuals) where the extension over multiple copies allows for several different physical states or network configurations. In particular, we are interested in epistemic uncertainty.

The “Deep Ensembles” technique consists of training an ensemble of deep neural networks on the same data with random initialization of each of the neural networks in the ensemble. By running all neural networks aggregating their prediction, one obtains the best in class uncertainty estimation at the cost of computational investment.

 

A “mask” is, in deep learning technology, a way to drop/hide artificial neurons, thus generating several slightly different model architectures, allowing a single model to mimic ensemble behaviour.

 

Masksembles can generate a range of models within which MC-Dropout and Deep Ensembles are extreme cases. It joins MC-Dropout’s light computational overhead and Deep Ensembles’ performance. Using many masks approximates MC-Dropout while using a set of non-overlapping, completely disjoint masks, yields an Ensemble-like behaviour.

 

Figure 3: The prediction task here is to classify the red vs blue data points drawn in the [−5, 5] range from two sinusoidal functions with added Gaussian noise. The background colour depicts the entropy assigned by different models to individual points on the grid, low entropy represented by blue and high entropy represented in yellow. The figure ranges from (a) Single model to (b) – (e) Masksembles models with different parameters and finally (f) – Ensembles model. For high mask-overlap values, Masksembles behaves almost like a Single Model, compare (a) and (b) but starts acting more and more like Ensembles as the mask-overlap value decreases, compare (e) and (f).

 

Conclusion

Neural Concept is already providing the possibility, even for non-specialists, to have confidence levels for predictions. The type of advanced research exposed in the last section will bring further benefits to engineers in terms of high performance and low computational costs.

Thus, we will continue to sustain engineers facing an always important question – «am I  sure about my predictions»?

 

Bibliography

Nikita Durasov, Timur Bagautdinov, Pierre Baqué, Pascal Fua (Computer Vision Laboratory, EPFL and Neural Concept): «Masksembles for Uncertainty Estimation», arXiv:2012.08334v1 [cs.LG], 15 Dec 2020

 

©Anthony Massobrio, Senior Technology Evangelist @ Neural Concept

The post The importance of uncertainty quantification for Deep Learning models in CAE appeared first on Neural Concept.

]]>
Deep Learning-based models for fast evaluations of airfoil performances in transonic flows https://neuralconcept.com/customer-cases/deep-learning-based-models-for-fast-evaluations-of-airfoil-performances-in-transonic-flows/ Mon, 18 Jan 2021 14:30:43 +0000 https://neuralconcept.com/?p=3958 ISAE-Supaero, a world leader in aerospace engineering higher education in France, and Neural Concept, have developed a new approach to high Mach aircraft design. From an incompressible flow simulation, Neural Concept’s Neural Networks are able to predict flow simulations at a high Mach number. Therefore, a full description of the simulations for the flight envelope […]

The post Deep Learning-based models for fast evaluations of airfoil performances in transonic flows appeared first on Neural Concept.

]]>
ISAE-Supaero, a world leader in aerospace engineering higher education in France, and Neural Concept, have developed a new approach to high Mach aircraft design. From an incompressible flow simulation, Neural Concept’s Neural Networks are able to predict flow simulations at a high Mach number. Therefore, a full description of the simulations for the flight envelope (Mach-Alpha), at high Mach, can be obtained accurately by running a single inexpensive simulation.

 

This methodology has been integrated into NCS and deployed into a NCS/Lite app, currently at demo stage:

 

High-Mach simulations are expensive and necessary for Aircraft design

 

For low-speed regimes, the flow over an airfoil can be considered as incompressible, thus simplifying the underlying physics, and making it simpler to study, analyze and simulate. However, when reaching higher speeds, the flow enters a transonic regime. Therefore, the simplifying assumption is not valid anymore as compressibility effects start to influence the overall performance of the airfoil, such as the formation of shock waves. The physics underlying these phenomena become then much more challenging to model.

 

shockwave around a fighter jet

Figure 1: Formation of a shockwave around a fighter jet

 

Hence, it is critical for the engineers to be able to understand the influence of these compressibility effects, at various operating conditions (flow velocity, angle of attack) to ensure reliability and uniformity in the performances of the design. To do so, it is normally necessary for Aerospace engineers to run many CFD simulations for various boundary conditions to cover the whole space of operating conditions. For transonic flows, with highly non-linear behavior, these simulations are very time-consuming and expensive, creating a bottleneck in the evaluation process.

 

Deep-Learning can predict many high-mach simulations from a single, non-expensive one

 

NCS high-mach simulations

 

The proposed methodology is based on Neural Concept’s Deep Learning models, which are trained to “transpose” simulations from low-mach to high-mach. It makes it possible to predict the whole Mach envelope, i.e., pressure fields at the various angle of attack and higher Mach numbers in a few seconds.  It uses as input a computationally non-expensive single incompressible flow simulation.

 

In this setting, the user is able to upload an airfoil geometry, with the corresponding simulation at a low Mach number. The model then predicts the corresponding pressure profile at a higher Mach number which can be specified by the user. This allows for much faster iterations between designs, and a better understanding of the complex phenomena occurring above a given speed.

 

The results presented in the report show that Neural Concept’s Deep Learning-based models can accurately describe the highly non-linear regions on the airfoil, where large pressure gradients occur, making it a reliable and robust solution to model compressibility effects of transonic flows on airfoils.

 

This approach has proven more accurate than the standard methodology used by Neural Concept’s users, where only a geometry would be fed into the predictor. Adding here the output of an incompressible flow simulation as an input to the Neural Network, seems to improve its generalization power. This is easily done in NCS by exploiting the options to provide so-called Input-Fields to the Neural Network.

 


 

For more information, please read the abstract:

https://neuralconcept.com/wp-content/uploads/2020/12/PGnet.pdf

 

 

© ISAE-Supaero: M.Bauerheim, D.Costero, V. Chapin, N.Gourdain & Neural Concept : L.Zampieri, P.Baqué

The post Deep Learning-based models for fast evaluations of airfoil performances in transonic flows appeared first on Neural Concept.

]]>
NCS to enhance the design of turbomachines https://neuralconcept.com/customer-cases/ncs-to-enhance-the-design-of-turbomachines/ Fri, 08 Jan 2021 16:03:47 +0000 https://neuralconcept.com/?p=3927 Turbomachines are very complex assemblies, which need to operate efficiently under a wide range of operating conditions. Simulation-driven design is now a key driver in the industry, but some major bottlenecks remain, which limit the improvements that are possible. How can NCS overcomes these limitations, and radically change the design process of turbomachines ? You […]

The post NCS to enhance the design of turbomachines appeared first on Neural Concept.

]]>
Turbomachines are very complex assemblies, which need to operate efficiently under a wide range of operating conditions. Simulation-driven design is now a key driver in the industry, but some major bottlenecks remain, which limit the improvements that are possible. How can NCS overcomes these limitations, and radically change the design process of turbomachines ? You can check our current work, in collaboration with NUMECA International, allowing for quasi-real-time performance maps and design space exploration.

In this video, you can see how the user is able to navigate on the performance map, evaluating the behavior of the design on specific operating conditions, for different values and views (pressure field, velocity field, etc…) . Then, the user can upload a new geometry, and get the instantaneous predictions of the model on the whole range of operating conditions.

To see the whole presentation made by Dirk Wunsch from NUMECA International, you can click on the following link:

Link to full presentation

 

The post NCS to enhance the design of turbomachines appeared first on Neural Concept.

]]>
Training deep learning models efficiently on the cloud https://neuralconcept.com/research/storing-data-efficiently-on-the-cloud/ Fri, 11 Dec 2020 10:03:35 +0000 https://neuralconcept.com/?p=3837 1.Storing data efficiently on the cloud   With Neural Concept Shape, you can use 3D numerical simulations as input to train your deep learning models. If we take the example of aerodynamic simulations, these CFD simulations results are usually much larger files than images or text (a single result can reach several hundreds of GB). […]

The post Training deep learning models efficiently on the cloud appeared first on Neural Concept.

]]>

1.Storing data efficiently on the cloud

 

With Neural Concept Shape, you can use 3D numerical simulations as input to train your deep learning models. If we take the example of aerodynamic simulations, these CFD simulations results are usually much larger files than images or text (a single result can reach several hundreds of GB). Hence storing a large amount of them can become an issue in the long term, as it would require to regularly scale up the hardware infrastructures accordingly.

The engineer might then face difficulties to stream through the files, to make relevant analysis, and it becomes a real limitation and bottleneck in the usage of deep learning for engineering applications. At Neural Concept, we are aware of this issue, and we addressed it by evaluating different solutions over time.

Our initial set up was a NFS (Network File System) solution. It is very convenient because it allows to access files the same way we would access them on the local storage of a machine and can be used by several users in a team. However, this solution does not scale well with the accumulation of data, and the increasing number of experiments done by users simultaneously. Moreover, it can quickly become expensive.

As an alternative, we chose to store the data in a secure cloud environment, using FUSE libraries, allowing us to easily access data as if it was on a local computer, while benefiting from the powerful and flexible cloud architecture.  FUSE, which stands for Filesystem in Userspace, is an interface for Unix-like operating systems that lets users create their own file systems. Using a fuse library, it is possible to mount a cloud storage bucket onto the local filesystem, and then applications can access files in the cloud storage as if it was on a local file system. The user of Neural Concept Shape is now able to train models directly from a secure cloud storage, without any impact on the speed of computation, as it was benchmarked internally.

Figure 1: Comparison of training speed, we can see that we reach similar training speeds with data stored on NFS, or on secure cloud environments.

2.Improving the training speed of deep learning models

 

Over the past years, the performance of GPUs has drastically improved, and are widely used in various deep learning applications. They allow a very fast and efficient computing, having large memory size available. Hence, the most modern GPUs are now able to tackle complex physics-based deep-learning challenges and deal with (very) refined geometries.

With Neural Concept Shape, we use 3D simulation data to train our models, which can be very heavy files if the simulation is extremely detailed. Most people would then tend to think that the main bottleneck when dealing with such data is the GPU itself, but it is not always the case. The main reason of slow-down (which can be critical for some applications) is sometimes the streaming of the data to the GPU. Indeed, for large files, and especially when the data is being fetched over the network from cloud storage, this can result in a drastic slow-down of the training process. It can then become a real limitation in the usage of deep learning for engineering applications.

This is why we are using the cache functionality of tensorflow data API ( https://www.tensorflow.org/api_docs/python/tf/data/Dataset#cache). We are able to cache the dataset to a local SSD disk during the training process, allowing very efficient retrieval of the data. Caching data means that it can be very efficiently and rapidly accessed as it is stored locally. After the initial pass over the dataset and the first iterations, the dataset gets cached and subsequent iterations go much faster.

Figure 2: Comparison of training speed, with or without the usage of local ssd cache

In the graph, we see that after the initial 150 steps (after which data has been cached to the local SSD Disk), the training speed increases and remains steady when using the cache.This enables the engineers using Neural Concept Shape to perform various experiments very efficiently, even when dealing with a large dataset, or very complex simulations.

 

Saswata Chakravarty, Lead Software Engineer @ Neural Concept

The post Training deep learning models efficiently on the cloud appeared first on Neural Concept.

]]>
Collaboration between Neural Concept and PSA https://neuralconcept.com/customer-cases/collaboration_between_neural_concept_and_psa/ Mon, 26 Oct 2020 12:29:05 +0000 https://neuralconcept.com/?p=3534 After evaluating the performance of Neural Network models on benchmark test cases (see Fig.1), PSA decided to push the collaboration further towards a real-time predictive model for external aerodynamics, applicable to production-level 3D simulations.   With this new step, PSA will aim at accelerating design cycles, the time between the ideation of a new design […]

The post Collaboration between Neural Concept and PSA appeared first on Neural Concept.

]]>
After evaluating the performance of Neural Network models on benchmark test cases (see Fig.1), PSA decided to push the collaboration further towards a real-time predictive model for external aerodynamics, applicable to production-level 3D simulations.

 

With this new step, PSA will aim at accelerating design cycles, the time between the ideation of a new design and the start of production. Another target is to optimize the performance of the next generations of vehicles, including larger autonomy and greater passenger comfort.

 

Figure 1 In this benchmark case, we compared the accuracy of a Geometric CNN to a production-level Gaussian Process. The dataset was composed of 800 samples of geometries described with up to 22 parameters.

 

The benchmark study compared Geometric CNNs to Gaussian-Process based regression models, specifically tuned for production-level simulations. It proved that, even though the Geometric CNN does not have access to the parametric description and is therefore much more broadly applicable than the Gaussian Process, it can also outperform the standard methods by a large margin.

 

Neural Concept Shape is a high-end deep learning software, which understands 3D shapes (CAD), and learns how they interact with the laws of physics (CAE). It is able to emulate full-fledged simulators, giving predictions in approximately 30 ms versus minutes to hours (or even days) for classic simulators. In other terms, engineers can use Neural Concept Shape to explore, manually or automatically, an infinite amount of designs without calling back the resource-consuming, time-consuming simulator.

 

NCS is the link between designers and simulation experts in the company, reducing lengthy iterations between teams. This allows to dramatically accelerate R&D cycles, enhance product performances, and solve the most difficult engineering challenges.

The post Collaboration between Neural Concept and PSA appeared first on Neural Concept.

]]>
Miniswys uses Neural Concept Shape for the design optimization of customized ultrasonic actuators https://neuralconcept.com/customer-cases/miniswys-uses-neural-concept-shape-for-the-design-optimization-of-customized-ultrasonic-actuators/ Thu, 08 Oct 2020 12:32:02 +0000 https://neuralconcept.com/?p=3421 Miniswys is a swiss company developing ultrasonic piezo-electric actuators to achieve precise bidirectional movements, reaching very large strokes with low driving voltage in compact applications. In order to develop such a product, the engineers perform Finite element analysis simulations, to estimate the dynamic behavior of the design, by performing modal analysis. It is used to […]

The post Miniswys uses Neural Concept Shape for the design optimization of customized ultrasonic actuators appeared first on Neural Concept.

]]>
Miniswys is a swiss company developing ultrasonic piezo-electric actuators to achieve precise bidirectional movements, reaching very large strokes with low driving voltage in compact applications. In order to develop such a product, the engineers perform Finite element analysis simulations, to estimate the dynamic behavior of the design, by performing modal analysis. It is used to estimate the resonance frequencies and structural modes of the geometry under various conditions.

 

However, each application has its own set of requirements, where small design variations can lead to completely different modal behavior. In this context, Miniswys and Neural Concept have been successfully collaborating over the past months, to build a 3D Deep Learning based surrogate model. It allows to get an instantaneous and precise estimation of the dynamic behavior of these actuators, based on geometric and/or boundary conditions variations.

 

Using Neural Concept Shape, Miniswys is able to explore in a very fast manner many different designs iterations, without the need of going through the full-fledged simulator at every step. Ultimately Miniswys is able to explore extensively the space of designs, to find innovative geometries, outperforming the classic ones, while drastically reducing the costs and time of the research and development phase. After using the tool, Raphaël Hoesli, CTO of Miniswys and directly involved in the project, expressed his satisfaction in the following words:

 

“Neural Concept Shape enables us to be much more efficient to design products meeting our customers’ requirements. The feedback from our design iterations is so fast that Miniswys’ engineers can see the evolution of the performance quasi instantaneously while changing the design parameters. In other words, slow iterations are replaced by quick predictions which give us the possibility to intuitively improve the performances of our actuators.”

 

These successful results encouraged Miniswys to continue using Neural Concept Shape to leverage on this surrogate model in shape optimization for piezo actuators.

Example of Miniswys linear ultrasonic actuator  Figure 1: Example of Miniswys linear ultrasonic actuator.

Figure 2: Comparison between the simulation and the neural network prediction on a test sample.

Neural Concept Shape is a high-end deep learning software, which understands 3D shapes (CAD), and learns how they interact with the laws of physics (CAE). It is able to emulate full-fledged simulators, giving predictions in approximately 30 ms versus minutes to hours (or even days) for classic simulators. In other terms, engineers can use Neural Concept Shape to explore, manually or automatically, an infinite amount of designs without calling back the resource-consuming, time-consuming simulator. This allows to dramatically accelerate R&D cycles, enhance product performances, and solve the most difficult engineering challenges.

 

The post Miniswys uses Neural Concept Shape for the design optimization of customized ultrasonic actuators appeared first on Neural Concept.

]]>
Neural Concept SA and EPFL CVLab at NeurIPS 2020: “MeshSDF: Differentiable Iso-Surface Extraction.” https://neuralconcept.com/research/meshsdf/ Tue, 06 Oct 2020 12:20:04 +0000 https://neuralconcept.com/?p=3454 A joint paper of Neural Concept SA and EPFL CVLab will be presented as a Spotlight presentation at Presentation for Neural Information Processing Systems Conference (NeurIPS) 2020 to reach a broader public to talk about MeshSDF that makes Deep SDF-based 3D generative models (VAEs) differentiable.   During the session, the team will introduce a differentiable […]

The post Neural Concept SA and EPFL CVLab at NeurIPS 2020: “MeshSDF: Differentiable Iso-Surface Extraction.” appeared first on Neural Concept.

]]>
A joint paper of Neural Concept SA and EPFL CVLab will be presented as a Spotlight presentation at Presentation for Neural Information Processing Systems Conference (NeurIPS) 2020 to reach a broader public to talk about MeshSDF that makes Deep SDF-based 3D generative models (VAEs) differentiable.

 

During the session, the team will introduce a differentiable way to produce explicit surface mesh representations from Deep Signed Distance Functions by removing the limitation of the Marching Cubes algorithm. The key insight is that by reasoning differentiate the 3D location of surface samples with respect to the underlying deep implicit field. The team exploit this to define MeshSDF, an end-to-end differentiable mesh representation which can vary its topology. They use two different applications to validate their theoretical insight: Single-View Reconstruction via Differentiable Rendering and Physically-Driven Shape Optimization.

 

The Neural Information Processing Systems Foundation is a non-profit corporation whose purpose is to foster the exchange of research on neural information processing systems in their biological, technological, mathematical and theoretical aspects. The Conference on Neural Information Processing Systems is the main venue where the most groundbreaking scientific publications in machine learning are published every year, with more than 10,000 attendees.

 

NIPS 2020 is held Sun 6th December through Sat the 12th, 2020 at Virtual-only.

 

Read the full paper on : https://arxiv.org/abs/2006.03997

The post Neural Concept SA and EPFL CVLab at NeurIPS 2020: “MeshSDF: Differentiable Iso-Surface Extraction.” appeared first on Neural Concept.

]]>
NUMECA uses Neural Concept’s Deep Learning platform to study the nature of turbulence https://neuralconcept.com/customer-cases/numeca_uses_neural_concept_deep_learning_platform_to_study_the_nature_of_turbulence/ Thu, 01 Oct 2020 16:56:17 +0000 https://neuralconcept.com/?p=3364 The most significant challenge in all areas of applied fluid mechanics is posed by the lack of understanding and thus poor prediction capability of turbulence dependent features. This leads to limited industrial confidence in CFD for many aeronautical applications such as flow detachment over an aircraft wing or shock-boundary layer interactions. Against this background, the […]

The post NUMECA uses Neural Concept’s Deep Learning platform to study the nature of turbulence appeared first on Neural Concept.

]]>
The most significant challenge in all areas of applied fluid mechanics is posed by the lack of understanding and thus poor prediction capability of turbulence dependent features. This leads to limited industrial confidence in CFD for many aeronautical applications such as flow detachment over an aircraft wing or shock-boundary layer interactions. Against this background, the HiFi-TURB project, which is coordinated by NUMECA, sets out a highly ambitious and innovative work programme to address influential deficiencies in turbulence modelling.

The large scale availability of High-Performance Computing (HPC) opens the door to a truly novel approach to turbulence model development. Exploiting Artificial Intelligence (AI) and Machine Learning (ML) techniques applied to a database of high-fidelity, scale-resolving simulations of test cases that contain most features of separated flow regions or complex 3D flows. Figure 1 shows an example of a flow field that is used as the basis for the turbulence modelling task.

Flow structure of T161 cascade

Figure 1: Flow structure of T161 cascade. Typical flow field used as input for the turbulence model improvement

The huge amount of data generated in these simulations requires a new approach to data mining. This is where Neural Concepts brings in its tool chain based on Deep-Learning, to analyse very large amounts of data provided by 3D scale resolving simulations.

Using Neural Concept’s Geometry-based Variational Auto-Encoders (VAE), NUMECA was able to gain first insights into correlations between tens of statistically averaged flow variables. The VAE compresses the data first in a physically meaningful way into so-called ‘embeddings’ and then reconstructs the original input from the compressed data. This is done to a very high accuracy, which allows to use the ML model as a replacement, a so-called surrogate, for the original data. The advantages are a much easier handling of the data, and the possibility of exploiting data mining and analysis techniques that help to understand the physics in the data.

Figure 2 shows an example of the possible analysis. The colors of the symbols on the 2D plot correspond to the value of the ‘embedding’ and are the same in the 3D view (left) and in the 2D plot (right). Points of the same color have the same value for all the considered physical quantities and the 3D view colored by the embedding value, gives us one global statistical representation for several physical quantities over the investigated domain. Both plots together provide a new perspective on the flow behaviour via the machine learning model. Figure 2, shows snapshots of the views used in the Graphical User Interface.

Structure found by the ML model and Statistical analysis of quantities

Figure 2: (top) Structure found by the ML model. (bottom) Statistical analysis of quantities

The post NUMECA uses Neural Concept’s Deep Learning platform to study the nature of turbulence appeared first on Neural Concept.

]]>
Collaboration of Neural Concept and Bosch on successful applications of 3D Deep Learning based surrogate models https://neuralconcept.com/customer-cases/collaboration-of-neural-concept-and-bosch-on-successful-applications-of-3d-deep-learning-based-surrogate-models/ Tue, 29 Sep 2020 13:18:34 +0000 https://neuralconcept.com/?p=3324 Neural Concept and Bosch Research collaborated over the past months on a set of successful applications of 3D geometric deep learning techniques powered by Neural Concept Shape (NCS) software.   More particularly, we achieved promising results on E-Drive motor housing simulations. Bosch Research engineers trained a deep Geometric Convolutional Neural Network (GCNN) to emulate accurately, […]

The post Collaboration of Neural Concept and Bosch on successful applications of 3D Deep Learning based surrogate models appeared first on Neural Concept.

]]>
Neural Concept and Bosch Research collaborated over the past months on a set of successful applications of 3D geometric deep learning techniques powered by Neural Concept Shape (NCS) software.

 

More particularly, we achieved promising results on E-Drive motor housing simulations. Bosch Research engineers trained a deep Geometric Convolutional Neural Network (GCNN) to emulate accurately, in a few ms, the fully fledged Finite Element software.

 

These successful results encouraged Bosch Research to continue the collaboration with Neural Concept on a further application of shape design optimization.

 

Comparison between the simulation and the neural network prediction on a test sample

 

“For the considered application, NCS performs clearly better than currently used surrogate models and therefore we see the potential of NCS for more use-cases.”

Roland Schirrmacher, Structural Dynamics and Acoustics engineer at Bosch

 

Neural Concept Shape is a high-end deep learning software, which understands 3D shapes (CAD) and learns how they interact with the laws of physics (CAE). It is able to emulate full-fledged simulators, giving predictions in approximately 30ms, versus minutes to hours (or even days) for classic simulators. In other terms, engineers can use Neural Concept Shape to explore, manually or automatically, an infinite amount of designs without calling back the resource consuming, time-consuming simulator. This allows to dramatically accelerated R&D cycles, enhance product performances and solve the most difficult engineering challenges.

 

The post Collaboration of Neural Concept and Bosch on successful applications of 3D Deep Learning based surrogate models appeared first on Neural Concept.

]]>