Physics-based Numerical Simulations Research Articles

Investigating worldwide strong motion databases to derive a collection of free-field records to select design-compatible waveforms for Switzerland

The process of choosing ground motions typically relies on assembling a collection of ground motions that match a desired spectrum. This selection process is guided by specific seismological criteria, including factors like earthquake magnitude, distance from the epicenter, site soil type, and the range of spectral periods that need to fit with the target spectrum. The selection algorithm and the available dataset of waveforms obviously play significant roles in this process. In many engineering and site response applications, it is essential that the input ground motion is representative for the shaking at the free surface of the Earth, and at times also a specific soil type may be required. However, real waveform databases often lack sufficient and/or consistent metadata related to the installation type and soil characterization of recording stations, as well as to the earthquake seismological parameters. This deficiency can lead to the selection of inappropriate waveforms, such as those recorded by stations situated within manmade structures (buildings, bridges, dams) or on a soil type different than the intended one. To address this issue, our approach for creating an appropriate waveform database applicable to Switzerland starts with the computation of seismic hazard disaggregation for return periods of 475 and 975 years. This computation helps identifying the magnitude-distance scenarios most relevant for the five seismic hazard zones defined in the Swiss building code. Once these magnitude-distance ranges are identified, we adhere to established standards regarding the quality control of three-component waveforms and their associated metadata. We assemble a database of waveforms by collating and homogenizing data from available global databases. In the interest of comprehensiveness, we also incorporate data obtained from 3D physics-based numerical simulations of strong-motion near the seismic source. Finally, we employ an algorithm that integrates the Eurocode 8 waveform selection criteria. This algorithm allows us to select and scale waveforms suitable for microzonation and structural analysis studies within each of Switzerland’s five seismic hazard zones. Selecting waveforms compatible with the target design spectra proves to be challenging due to the stringent criteria imposed by Eurocode 8. This challenge arises from the scarcity of recorded waveforms with verified metadata and precise site characterization in the desired magnitude-distance ranges.

Predicting the seismic ground-motion parameters: 3D physics-based numerical simulations combined with artificial neural networks

Typically, it is challenging to incorporate near-surface soils into 3D physics-based numerical simulations (PBSs) for ground-motion prediction. The low shear wave speed of near-surface soils, coupled with the complexity of the soil seismic response, poses significant difficulties. To overcome these limitations, a hybrid approach was proposed in this study, combining PBSs with artificial neural networks (ANNs). The essence of the hybrid method can be summarized as follows: (1) development of ANN models, establishing a strong-motion database, training the ANNs on it to predict the ground-motion parameters for East–West (EW), North–South (NS), and Vertical (UD) directions afterward; (2) establishment of 3D PBS model, obtaining the ground-motion parameters of the bedrock face corresponding to a certain shear wave speed; (3) application of the trained ANNs to predict the ground-motion parameters on the ground surface, taking the simulated results and related site parameters as inputs, and the outputs are peak ground acceleration (PGA) and 5% damped spectral accelerations (Sa) at different periods on the ground surface. In this study, ANN models were trained on a strong-motion database based on Kiban–Kyoshin Network (KiK-net). After several verifications of the ANN predictions, a case study of the 21 October 2016 Mw6.2 Central Tottori earthquake was conducted. In addition to the comparison with observations, the broadband (0.1–10 Hz) results of the hybrid method were also compared with the results that obtained by transfer function based on recorded data and Next Generation Attenuation (NGA)-West2 ground-motion prediction equations (GMPEs) to demonstrate the effectiveness and applicability of the proposed method. In addition, the distribution of Sa for four periods in simulated area was presented. The performance of the hybrid method for predicting broadband ground-motion characteristics was generally satisfactory.

Early Detection of Extreme Storm Tide Events Using Multimodal Data Processing

Sea-level rise is a well-known consequence of climate change. Several studies have estimated the social and economic impact of the increase in extreme flooding. An efficient way to mitigate its consequences is the development of a flood alert and prediction system, based on high-resolution numerical models and robust sensing networks. However, current models use various simplifying assumptions that compromise accuracy to ensure solvability within a reasonable timeframe, hindering more regular and cost-effective forecasts for various locations along the shoreline. To address these issues, this work proposes a hybrid model for multimodal data processing that combines physics-based numerical simulations, data obtained from a network of sensors, and satellite images to provide refined wave and sea-surface height forecasts, with real results obtained in a critical location within the Port of Santos (the largest port in Latin America). Our approach exhibits faster convergence than data-driven models while achieving more accurate predictions. Moreover, the model handles irregularly sampled time series and missing data without the need for complex preprocessing mechanisms or data imputation while keeping low computational costs through a combination of time encoding, recurrent and graph neural networks. Enabling raw sensor data to be easily combined with existing physics-based models opens up new possibilities for accurate extreme storm tide events forecast systems that enhance community safety and aid policymakers in their decision-making processes.

Efficient AI-Physics hybrid model with productive capabilities to reduce the time of history matching and scenario assessment; a case study: Minagish oil field

This paper proposes a novel approach that combines physics-based numerical simulation with deep-learning neural networks to create an AI-Physics hybrid model for reservoir simulation. Our primary objective is to reduce the time of history match and achieve the best match between model and observation data using our hybrid model. We trained our model using historical data and created a reliable forecasting model that predicts field behavior and proposes new scenarios to improve production in the upcoming years. To test our model, we combined AI physics history training with blind test prediction calculations of the remaining oil map, and the AI physics model created a forecast scenario definition based on this map. Our proposed simulation method can reduce the time of history matching and scenario evaluation by 90-95%. We created three improved forecast scenarios based on predefined scenarios that can produce millions of standard barrels more oil over three years than the original development scenario. The results demonstrate the potential of our AI-Physics hybrid model in revolutionizing reservoir simulation in the oil and gas industry. Our approach can significantly reduce simulation time, improve forecasting accuracy, and optimize production strategies, leading to significant economic benefits.

Hybrid data-driven and physics-based simulation technique for seismic analysis of steel structural systems

This paper proposes 1) a new hybrid analysis technique by integrating a data-driven method with a physics-based technique to perform nonlinear analysis of steel structural systems under seismic loading, 2) two component-based data-driven models (PI-SINDy and DPI-SINDy) for predicting the nonlinear hysteretic response of steel seismic fuses with and without hysteretic degradation. The proposed hybrid data-driven and physics-based simulation (HyDPS) technique offers an efficient approach for seismic analysis of structures and is expected to address the challenges associated with computational cost and modeling uncertainties inherent in physics-based numerical simulations. In this technique, the well-understood components of the structure modeled numerically are combined with the critical components of the structure simulated using one of the data-driven models developed in this study. The proposed data-driven models were trained using experimental and numerical hysteresis data. The results show that these data-driven models can accurately and efficiently predict the nonlinear hysteretic response of steel structural components with and without degradation. Furthermore, the performance of the HyDPS technique powered by the PI-SINDy model is verified in the presence of noise using response history analyses performed on a steel buckling-restrained braced frame.

A new surge index with the incorporation of cyclone track approach angle information for the Bay of Bengal

A novel tropical cyclone surge potential index (TCSPI) for estimating tropical cyclone (TC) induced potential peak surge levels near the landfall locations for the entire coast of the Bay of Bengal (BoB) is presented. The proposed TCSPI incorporates parameters that are often readily available in real-time for BoB: maximum sustained wind speed (Vmax), the radius of given wind (R), translation speed (S), approach angle (θ), coastal geometry (a, b), and bathymetry information (L30). The inclusion of approach angle and associated coastal geometry information led to improvements of the TCSPI to other existing surge indices. A retrospective analysis of TCSPI using the Indian Meteorological Department (IMD) and Joint Typhoon Warning Center (JTWC) tropical cyclone (TC) best track data from 1990 to 2021 for BoB suggests that the index captures historical events reasonably well. In particular, it explains up to ∼90% of observed surge variance and up to ∼92% of those peak surge values obtained from physics-based numerical simulation using the Advanced Circulation (ADCIRC) model. Further, we utilized renowned regression-based supervised machine learning (ML) models on small data samples. The Linear Regression (LR), Regression Tree (TR), Support Vector Machine (SVM), Ensembles of Trees (EnTR), and Gaussian Process Regression (GPR) models are incorporated that can rapidly predict peak surge height across an extensive coastal region of BoB, using the same input parameter information (i.e., Vmax, R, S,L30, and θ) that have been considered in the TCSPI at the landfall time frame. Finally, the observed peak surge values associated with historical TC events in the BoB are compared with those estimated by TCSPI, ADCIRC, and ML models. The errors associated with TCSPI are comparatively less. The consideration of approach angle information in the TCSPI has improved the accuracy of the overall prediction of peak surge height associated with historical TCs by up to 17% and 19% for observed and ADCIRC-simulated peak surges, respectively. The development of such a prediction methodology can reduce the numerical computation requirements to improve the surge hazard maps, and hence, proper implementation based on surge risk is possible.

Engineering validation of BB-SPEEDset, a data set of near-source physics-based simulated accelerograms

Physics-based numerical simulations (PBS) have progressed to the level of providing a realistic description of earthquake ground motion and of its variability, both in time and space, thus enabling to fill the knowledge gaps due to the sparsity of recordings (especially in the near-source region of strong earthquakes). Nevertheless, to build confidence in the utilization of the PBS by the engineering community, simulated accelerograms need to be validated against recorded data from both seismological and engineering perspectives. This article aims at extending the validation of BB-SPEEDset, a data set of near-source broadband simulated accelerograms from multiple regions and faulting styles, obtained by the spectral element computer code SPEED. In addition to seismological checks of BB-SPEEDset proving the absence of systematic biases with respect to a near-source records data set, in this work, the validation is addressed in terms of engineering demand parameters (EDPs) of elastoplastic single-degree-of-freedom systems, taking advantage of a ground motion selection tool including simulated accelerograms. It is found that, when simulated and recorded accelerograms are selected according to the same spectral compatibility criteria, consistent statistical distributions of EDPs are obtained from the two sets. To highlight the potentialities of BB-SPEEDset for near-source analyses, an example of utilization of spectrum-compatible pulse-like motions for structural inelastic analyses is also given, resulting in a good agreement with literature solutions in terms of inelastic displacement demands.

Data-driven production optimization using particle swarm algorithm based on the ensemble-learning proxy model

Production optimization is of significance for carbonate reservoirs, directly affecting the sustainability and profitability of reservoir development. Traditional physics-based numerical simulations suffer from insufficient calculation accuracy and excessive time consumption when performing production optimization. We establish an ensemble proxy-model-assisted optimization framework combining the Bayesian random forest (BRF) with the particle swarm optimization algorithm (PSO). The BRF method is implemented to construct a proxy model of the injection–production system that can accurately predict the dynamic parameters of producers based on injection data and production measures. With the help of proxy model, PSO is applied to search the optimal injection pattern integrating Pareto front analysis. After experimental testing, the proxy model not only boasts higher prediction accuracy compared to deep learning, but it also requires 8 times less time for training. In addition, the injection mode adjusted by the PSO algorithm can effectively reduce the gas–oil ratio and increase the oil production by more than 10% for carbonate reservoirs. The proposed proxy-model-assisted optimization protocol brings new perspectives on the multi-objective optimization problems in the petroleum industry, which can provide more options for the project decision-makers to balance the oil production and the gas–oil ratio considering physical and operational constraints.

TCAD Device Modeling and Simulation Study of Organic Field Effect Transistor-Based pH Sensor with Tunable Sensitivity for Surpassing Nernst Limit

A dual-gate organic field effect transistor (DG-OFET)-based pH sensor is proposed that will be able to detect the variations in the aqueous (electrolyte) medium. In this structure, a source-sided underlap technique with a dual-gate sensing approach has been used. The change in ON-current (ION) was observed due to parallel examination of electrolytes in two gates underlapping the region of the structure. For the evaluation of the sensitivity of DG-OFET, the change in the drain current was exploited for different pH and corresponding charge densities utilizing 2D physics-based numerical simulation. The simulation results were extracted with the help of the software package Silvaco TCAD-ATLAS. The simulated results display that the proposed DG-OFET shows significantly higher sensitivity for high-k dielectrics. The voltage sensitivity achieved by DG-OFET with SiO2 as a dielectric in our work is 217.53 mV/pH which surpasses the Nernst Limit nearly four times. However, using a high-k dielectric (Ta2O5) increases it further to 555.284 mV/pH which is more than nine times the Nernst Limit. The DG-OFET pH sensor has a lot of potential in the future for various flexible sensing applications due to its flexibility, being highly sensitive, biocompatible and low-cost.

Hybrid digital twin for conditional lubricant oil transport simulation and oil consumption prediction in internal combustion engines

Reducing lubricant oil consumption (LOC) has long been an interest to engine manufacturers in an effort to reduce overall emissions from internal combustion (IC) engines. However, directly correlating engine operating conditions with LOC rates is a challenging task given its complex and multi-physics nature. Building on previous experimental observations and modeling efforts, a baseline hybrid model, which combines traditional physics-based numerical simulation with modern machine learning techniques, was developed. As the first step, the model is to compute the oil exiting the top ring gap with given oil supply rate and locations to the second land. Studies performed using the model reveal a strong dependency of LOC on engine speed, load as well as the relative locations of piston ring gaps. In particular, the close proximity of the oil source on second land to the top ring gap and the reverse gas flow at low load conditions can both induce high LOC. Piston ring gap rotations and the different oil supply mechanisms to the piston ring pack are also found to heavily influence the oil consumption behavior and oil film thickness distribution pattern. The present work is the first step to a more comprehensive model for predicting LOC for healthy systems of piston ring packs. The efficiency of the model made it possible to study the oil redistribution that takes many engine cycles to reach equilibrium and rotation of the ring gaps that could happen in reality at a time scale of minutes and even longer.

Regional physics-based simulation of ground motion within the Rhȏne Valley, France, during the MW 4.9 2019 Le Teil earthquake

In this paper we introduce the 3D physics-based numerical simulations (PBS) of ground motion during the Nov 11, 2019, MW 4.9 Le Teil earthquake, which occurred in a low-to-moderate seismicity area in the South-East of France, within the Rhône river valley, which hosts several operating nuclear installations. The numerical code SPEED, developed at Politecnico di Milano, Italy, was used to produce the PBS. After introducing the criteria to construct the numerical model, based on the relatively limited data available, a numerical convergence test was made to identify the frequency range for accurate simulations. Furthermore, the performance of the numerical results against the available strong motion records was assessed quantitatively using Goodness-of-fit (GoF) measures. According to the GoF scores, a good-to-excellent agreement was found on the horizontal components up to 8 Hz, showing that, even without a very detailed 3D numerical model of the medium, that would imply detailed knowledge of the basin shape, of the bedrock-to-basin impedance ratio, as well as of the damping ratio in the basin and its dependence on frequency, the PBS may provide realistic broadband predictions of earthquake ground motion. Nevertheless, as shown by the poorer performance on the vertical component, the high-frequency limitations of PBS, also in relation to the energy radiated by the kinematic source model, is still an issue to be carefully addressed. In spite of these limitations, the results obtained in this work demonstrate that PBS, if suitably calibrated and validated, can be either an alternative or a useful complement to empirical ground motion models, especially in those cases where the region- and site-specific features of ground shaking, including near-source conditions, are typically not accounted for by ergodic empirical models, such as for the seismic risk evaluation of large urban areas and/or of strategic structures, infrastructures and industrial plants.

Improving deep learning performance for predicting large-scale geological {{CO}_{2}} sequestration modeling through feature coarsening

Physics-based reservoir simulation for fluid flow in porous media is a numerical simulation method to predict the temporal-spatial patterns of state variables (e.g. pressure p) in porous media, and usually requires prohibitively high computational expense due to its non-linearity and the large number of degrees of freedom (DoF). This work describes a deep learning (DL) workflow to predict the pressure evolution as fluid flows in large-scale 3-dimensional(3D) heterogeneous porous media. In particular, we develop an efficient feature coarsening technique to extract the most representative information and perform the training and prediction of DL at the coarse scale, and further recover the resolution at the fine scale by spatial interpolation. We validate the DL approach to predict pressure field against physics-based simulation data for a field-scale 3D geologic CO_2 sequestration reservoir model. We evaluate the impact of feature coarsening on DL performance, and observe that the feature coarsening not only decreases the training time by >74% and reduces the memory consumption by >75%, but also maintains temporal error 0.63% on average. Besides, the DL workflow provides predictive efficiency with 1406 times speedup compared to physics-based numerical simulation. The key findings from this research significantly improve the training and prediction efficiency of deep learning model to deal with large-scale heterogeneous reservoir models, and thus it can also be further applied to accelerate workflows of history matching and reservoir optimization for close-loop reservoir management.

Regional-scale assessment of earthquake-induced slope displacement considering uncertainties in subsurface soils and hydrogeological condition

Realistic prediction of coseismic landslides is crucial for the design of civil infrastructures and geotechnical systems in seismically active regions. Coseismic landslides are complicated nonlinear and large-deformation processes triggered by ground shaking. To date, many gaps still remain in the scientific understanding of the landslides, particularly in areas lacking information on the subsurface stratigraphy. Analytical methods for estimating coseismic landslides have been based on highly simplified models, where many key influential factors have been overlooked, leading to large uncertainty in using empirical models for landslide prediction. In addition, more evidence recently underscores the necessity of modeling coupled effects between topography and soil amplification, which gives rise to complex wave propagation patterns due to scattering and diffracting of waves within the low-velocity near-surface layers. Neglecting topographic amplification and soil layers could result in significantly underpredicted regional-scale hazards of landslides. In this study, a regional-scale coseismic landslide simulation is carried out using the integral Spectral Element Method (SEM)-Newmark method for Hong Kong Island to quantify the coupled effect of 3D topography, subsurface soil condition as well as hydrogeological conditions. By leveraging extensive geological borehole data, a rigorous statistical model that characterizes uncertainties associated with subsurface soils is established. Influence of key factors that are not well considered in previous coseismic landslide studies, such as variation in irregular rockhead and spatially distributed shear-wave velocity are quantitatively analyzed. Numerical results demonstrate that by rigorously incorporating multiple sources of subsurface soil uncertainties, the total area of landslides in the study region can increase up to 30% compared with benchmark model as evidenced in the physics-based numerical simulation. These case studies will provide a valuable opportunity to revisit key factors that influence coseismic landslides using measured field data.

Validation of physics-based ground shaking scenarios for empirical fragility studies: the case of the 2009 L’Aquila earthquake

This paper explores and validates the use of ground shaking scenarios generated via 3D physics-based numerical simulations (PBS) for seismic fragility studies. The 2009 L’Aquila seismic event is selected as case-study application, given the availability of a comprehensive post-earthquake database, gathering observed seismic damages detected on several building typologies representative of the Italian built environment, and of a validated numerical model for the PBS of ground shaking scenarios. Empirical fragility curves are derived as a function of different seismic intensity measures, by taking advantage of an improved statistical technique, overcoming possible uncertainties in the resulting estimates entailed by data aggregation. PBS-based fragility functions are compared to the corresponding sets of curves relying on updated ShakeMaps. The predictive capability of the adopted simulation strategies is then verified in terms of seismic damage scenarios, by respectively coupling PBS- and ShakeMap-based fragility models with the corresponding ground shaking scenarios. Comparison of observed and predicted damage distributions highlights the suitability of PBS for region-specific seismic vulnerability and risk applications.

Coupled thermo-mechanical numerical model to minimize risk in large-format additive manufacturing of thermoplastic composite designs

The collapse of deposited thermoplastic composite material under self-weight presents a risk in large-format extrusion-based additive manufacturing. Two critical processing parameters, extrusion temperature and deposition rate, govern whether a deposited layer is stable and bonds properly with the previously deposited layer. Currently, the critical parameters are determined via a trial-and-error approach. This research work uses a simplified physics-based numerical simulation to determine a suitable combination of the parameters that will avoid the collapse of the deposited layer under self-weight. The suitability of the processing parameters is determined based on the maximum plastic viscous strains computed using a sequentially coupled thermo-mechanical numerical model. This computational tool can efficiently check if a combination of temperature and extrusion rate causes layer collapse due to self-weight, and hence minimize the manufacturing risk of large-format 3D-printed parts.

Blind broad-band (0–10 Hz) numerical prediction of the 3-D near field seismic response of anMW6.0 extended fault scenario: application to the nuclear site of Cadarache (France)

SUMMARYIn this paper, physics-based numerical simulation (PBS) is employed to render a broad-band (0–10 Hz) realization of the near-field seismic response of the experimental nuclear site of Cadarache, located nearby the active Middle Durance Fault (southeastern France). The sensitivity of the earthquake numerical model to geological features is investigated by comparison with geophysical measurements and past aftershock and by highlighting the amplification induced by the soft sediments below Cadarache. The blind prediction of an MW6 target earthquake is approached by synthesizing four different finite-fault scenarios. The outcome is compared to the standard ground motion prediction equations (GMPEs), unveiling a possible GMPE overestimation of the pseudospectral acceleration ordinates at short natural periods, supporting the actual need to integrate synthetic and empirical predictions when direct observations are not available.

Self-trained artificial neural network for physical classification of ionospheric radar data

In the paper an algorithm for automatic classification of signals received by decameter coherent scatter radars is described. The algorithm is self-trained on unlabeled dataset and is based on training optimal physics-based classification using clusterization results. The approach allows to find optimal (hidden) classification, based on physically interpretable parameters of the received data, both obtained during physics-based numerical simulation and measured experimentally. As a result we trained the classification artificial neural network and found 11 well-interpretable classes from physical point of view in the available data. Other found classes are not interpretable from physical point of view, demonstrating the importance of taking into account radiowave propagation for correct classification. The advantages of the method include the possibility of its automatic training: it does not require preliminary data analysis by an expert, and it is self-trained to separate available data into the classes, that can be interpreted from a physical point of view.

A procedure for 3D simulation of seismic wave propagation considering source‐path‐site effects: Theory, verification and application

AbstractThis paper aims at obtaining a semi‐analytical and semi‐numerical 3D model of source‐to‐site seismic wave propagation due to kinematic finite‐fault sources. To this end, a two‐step procedure integrating the frequency‐wavenumber (FK) approach with the spectral element method (SEM) is proposed based on the concept of domain reduction. First, the broadband responses of a stratified crust are accurately calculated by using a novel FK approach and are converted into effective seismic inputs around the region of interest. After that, the seismic wavefields at local and regional scales arising from complex geological and topographical conditions are finely simulated using the SEM, and a perfectly matched layer absorbing boundary condition is simultaneously applied to realize the absorption of outgoing waves. In this procedure, a hybrid source modelling scheme that combines the low‐wavenumber deterministic and high‐wavenumber stochastic components on the fault plane is introduced, effectively addressing the high‐frequency motion radiated from the source rupture process. Subsequently, the proposed FK‐SEM procedure is verified step‐by‐step using the point source and finite‐fault source models. To illustrate the feasibility of the procedure, 3D physics‐based numerical simulations (PBSs) of two seismic events, including a historical Sanhe‐Pinggu earthquake and a well‐recorded Yangbi earthquake, are performed. The case studies validate that the proposed FK‐SEM procedure allows a significant reduction in computational effort and a substantial improvement in modelling resolution and can be applied to the source‐to‐site broadband synthetics of earthquake scenarios with limited resources. In addition, this coupled geophysics‐engineering simulation meets the requirements of time‐history analysis for engineering structures, which facilitates the study of soil‐structure interactions and regional‐scale building damage assessment.

A Probabilistic Physics-Based Seismic Hazard Model for the Alborz Region, Iran

ABSTRACTThe seismic activity rate is one of the most significant factors in seismic hazard modeling. Although it is usually estimated from observed seismicity, a complete picture of the possible earthquakes is not always available since catalogs of the observed earthquakes are short and incomplete. Long-term physics-based numerical simulations, providing a comprehensive range of earthquakes, are a decent way to overcome such deficiency. With this contribution, we built a seismic hazard model for the Alborz region, Iran, using a long-term physics-based synthetic earthquake catalog, enriched with the additional consideration of background seismicity derived from a deformation model. 200,000 yr synthetic catalogs for the Alborz region, Iran, are used and validated by considering the recurrence time of large-magnitude events estimated from the paleoseismological investigation on individual faults. The magnitude–frequency distribution (MFD) from the synthetic earthquake catalog is then compared with the MFD based on observation, which overall indicates good compatibility, although there are discrepancies for some faults. The estimated peak ground acceleration (PGA) for the Alborz region varies in the ranges of 0.16–0.52g and 0.27–1.0g for 10% and 2% probability of exceedance in 50 yr, respectively. The absolute natural logarithm differences averaged across the region are ∼0.21, corresponding to an average of 23% difference in PGA values in comparison with the most up-to-date observed-based hazard model. Hazard curves for several populated cities are also presented and compared with the other independent estimates. The proposed procedure could be an alternative approach to evaluate seismic hazard for a seismically active region, in particular for those without a complete catalog of observed earthquakes.

An approach for process optimisation of the Automated Fibre Placement (AFP) based thermoplastic composites manufacturing using Machine Learning, photonic sensing and thermo-mechanics modelling

The automated fibre placement (AFP) process is a complex manufacturing technique with many variables which affect the final part quality. Inverse Machine Learning (ML) models can be used as decision-aid tools for optimising thermoplastic composites manufacturing. However, a common challenge of ML application in manufacturing is the acquisition of relevant and sufficient data. To overcome this small-data learning problem, a hybrid approach has been proposed here which combines the benefits of ML algorithms such as the Artificial Neural Networks (ANN), virtual sample generation (VSG) methods, physics-based numerical simulations and data obtained from experiments and photonic sensors, to enhance the manufacturing process.