Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Randomized methods”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,495 records · Page 83
All sourcesNASA NTRSDOE OSTI
DOE OSTI2023

Moving Beyond Mortality: Development and Application of a Desirability of Outcome Ranking (DOOR) Endpoint for Hospital-Acquired Bacterial Pneumonia and Ventilator-Associated Bacterial Pneumonia

Abstract Background Hospital-acquired and ventilator-associated bacterial pneumonia (HABP/VABP) are frequently caused by multidrug-resistant organisms. Patient-centered endpoints in clinical trials are needed to develop new antibiotics for HABP/VABP. Desirability of outcome ranking (DOOR) is a paradigm for the design, analysis, and interpretation of clinical trials based on a patient-centered, benefit-risk evaluation. Methods A multidisciplinary committee created an infectious diseases DOOR endpoint customized for HABP/VABP, incorporating infectious complications, serious adverse events, and mortality. We applied this to 2 previously completed, large randomized controlled trials for HABP/VABP. ZEPHyR compared vancomycin to linezolid and VITAL compared linezolid to tedizolid. For each trial, we evaluated the DOOR distribution and probability, including DOOR component and partial credit analyses. We also applied DOOR in subgroup analyses. Results In both trials, the HABP/VABP DOOR demonstrated similar overall clinical outcomes between treatment groups. In ZEPHyR, the probability that a participant treated with linezolid would have a more desirable outcome than a participant treated with vancomycin was 50.2% (95% confidence interval [CI], 45.1%­−55.3%). In VITAL, the probability that a participant treated with tedizolid would have a more desirable outcome than a participant treated with linezolid was 48.7% (95% CI, 44.8%–52.6%). The DOOR component analysis revealed that participants treated with tedizolid had a less desirable outcome than those treated with linezolid when considering clinical response alone. However, participants with decreased renal function had improved overall outcomes with tedizolid. Conclusions The HABP/VABP DOOR provided more granular information about clinical outcomes than is typically presented in clinical trials. HABP/VABP trials would benefit from prospectively using DOOR.

Immunology↗
NASA NTRS1982

Three-dimensional stylus profilometry

The method of three-dimensional stylus profilometry is presented together with a computer-based system for transforming the surface topography of a specimen into an intensity image. Examples are given in which intensity images obtained by stylus profilometry are compared with scanning electron micrographs and two other representations of the stylus data. The comparisons are made for three types of surface irregularities: a discrete feature, a periodic profile surface, and a random profile surface. It is shown that stylus profilometry provides high vertical resolution for low sloped surfaces and thus can be useful in revealing features not easily detectable with other methods.

Teague, E. C.↗
DOE OSTI2026

Efficient many-jet event generation with flow matching

We apply for the first time, to the best of our knowledge, the flow matching method to the problem of phase-space sampling for event generation in high-energy collider physics. By training the model to remap the random numbers used to generate the momenta and helicities of the scattering matrix elements as implemented in the portable partonic event generator pepper, we find substantial efficiency improvements in the studied processes. We focus our study on the highest final-state multiplicities in Drell-Yan and top-antitop pair production used in simulated samples for the Large Hadron Collider, which computationally are the most relevant ones. We find that the unweighting efficiencies improve by factors of 184 and 25, respectively, when compared to the standard approach of using a vegas-based optimization. We also compare continuous normalizing flows trained with flow matching against the previously studied normalizing flows based on coupling layers and find that the former leads to better results, faster training and a better scaling behavior across the studied multiplicity range, while the latter evaluate faster. When combining the advantages of both methods using the regflow approach, we find parton-level unweighted event generation walltime gains of about a factor of 10 at the highest final-state multiplicities.

Bothmann, E. [CERN; Gottingen U.] (ORCID:000000016↗
NASA NTRS2008

Using LDPC Code Constraints to Aid Recovery of Symbol Timing

A method of utilizing information available in the constraints imposed by a low-density parity-check (LDPC) code has been proposed as a means of aiding the recovery of symbol timing in the reception of a binary-phase-shift-keying (BPSK) signal representing such a code in the presence of noise, timing error, and/or Doppler shift between the transmitter and the receiver. This method and the receiver architecture in which it would be implemented belong to a class of timing-recovery methods and corresponding receiver architectures characterized as pilotless in that they do not require transmission and reception of pilot signals. Acquisition and tracking of a signal of the type described above have traditionally been performed upstream of, and independently of, decoding and have typically involved utilization of a phase-locked loop (PLL). However, the LDPC decoding process, which is iterative, provides information that can be fed back to the timing-recovery receiver circuits to improve performance significantly over that attainable in the absence of such feedback. Prior methods of coupling LDPC decoding with timing recovery had focused on the use of output code words produced as the iterations progress. In contrast, in the present method, one exploits the information available from the metrics computed for the constraint nodes of an LDPC code during the decoding process. In addition, the method involves the use of a waveform model that captures, better than do the waveform models of the prior methods, distortions introduced by receiver timing errors and transmitter/ receiver motions. An LDPC code is commonly represented by use of a bipartite graph containing two sets of nodes. In the graph corresponding to an (n,k) code, the n variable nodes correspond to the code word symbols and the n-k constraint nodes represent the constraints that the code places on the variable nodes in order for them to form a valid code word. The decoding procedure involves iterative computation of values associated with these nodes. A constraint node represents a parity-check equation using a set of variable nodes as inputs. A valid decoded code word is obtained if all parity-check equations are satisfied. After each iteration, the metrics associated with each constraint node can be evaluated to determine the status of the associated parity check. Heretofore, normally, these metrics would be utilized only within the LDPC decoding process to assess whether or not variable nodes had converged to a codeword. In the present method, it is recognized that these metrics can be used to determine accuracy of the timing estimates used in acquiring the sampled data that constitute the input to the LDPC decoder. In fact, the number of constraints that are satisfied exhibits a peak near the optimal timing estimate. Coarse timing estimation (or first-stage estimation as described below) is found via a parametric search for this peak. The present method calls for a two-stage receiver architecture illustrated in the figure. The first stage would correct large time delays and frequency offsets; the second stage would track random walks and correct residual time and frequency offsets. In the first stage, constraint-node feedback from the LDPC decoder would be employed in a search algorithm in which the searches would be performed in successively narrower windows to find the correct time delay and/or frequency offset. The second stage would include a conventional first-order PLL with a decision-aided timing-error detector that would utilize, as its decision aid, decoded symbols from the LDPC decoder. The method has been tested by means of computational simulations in cases involving various timing and frequency errors. The results of the simulations ined in the ideal case of perfect timing in the receiver.

Jones, Christopher↗
NASA NTRS2010

A Model for Backscattering from Quasi Periodic Corn Canopies at L-Band

In this study, a model for backscattering at L-band from a corn canopy is proposed. The canopy consists of a quasi-periodic distribution of stalks and a random distribution of leaves. The Distorted Born Approximation (DBA) is employed to calculate the single scattered return from the corn field. The new feature of the method is that the coherence of the stalks in the row direction is incorporated in the model in a systematic fashion. Since the wavelength is on the order of the distance between corn stalks in a row, grating lobe behavior is observed at certain azimuth angles of incidence. The results are compared with experimental values measured in Huntsville, Alabama in 1998. The mean field and the effective dielectric constant of the canopy are obtained by using the Foldy approximation. The stalks are placed in the effective medium in a two dimensional lattice to simulate the row structure of a corn field. In order to mimic a real corn field, a quasi-periodic stalk distribution is assumed where the stalks are given small random perturbations about their lattice locations. Corn leaves are also embedded in the effective medium and the backscattered field from the stalks and the leaves is computed. The backscattering coefficient is calculated and averaged over successive stalk position perturbations. It is assumed that soil erosion has smoothed the soil sufficiently so that it can be assumed flat. Corn field backscatter data was collected from cornfields during the Huntsville 98 experimental campaign held at Alabama A&M University Research Station, Huntsville, Alabama in 1998 using the NASA/GW truck mounted radar. Extensive ground truth data was collected. This included soil moisture measurements and corn plant architectural data to be used in the model. In particular, the distances between the stalks in a single row have been measured. The L-band radar backscatter data was collected for both H and V polarizations and for look angles of 15o and 45o over a two week period under varying soil moisture conditions. These measured backscattering values will be compared with the model backscattering values and a discussion of the results will be presented.

Lang, R.↗
NASA NTRS2006

Free-field Calibration of the Pressure Sensitivity of Microphones at Frequencies up to 80 kHz

A free-field (FF) substitution method for calibrating the pressure sensitivity of microphones at frequencies up to 80 kHz is demonstrated with both grazing and normal incidence geometries. The substitution-based method, as opposed to a simultaneous method, avoids problems associated with the non-uniformity of the sound field and, as applied here, uses a 1/2 -inch air-condenser pressure microphone as a known reference. Best results were obtained with a centrifugal fan, which is used as a random, broadband sound source. A broadband source minimizes reflection-related interferences that often plague FF measurements. Calibrations were performed on 1/4-inch FF air-condenser, electret, and micro-electromechanical systems (MEMS) microphones in an anechoic chamber. The accuracy of this FF method is estimated by comparing the pressure sensitivity of an air-condenser microphone, as derived from the FF measurement, with that of an electrostatic actuator calibration and is typically 0.3 dB (95% confidence), over the range 2-80 kHz.

Herring, G. C.↗
DOE OSTI2020

An Evaluation of a Neutron Time Correlated Interrogation Method for Measurement of Fissile Content in Research Reactor Spent Fuel Assemblies

There are many research reactors worldwide that have been, or are being converted from the use of high enrichment uranium (HEU) to low enrichment uranium (LEU, < 20% enriched). The verification of fissile content and initial enrichment of the spent fuel is needed for the effective safeguards of the fuel. The advanced experimental fuel counter (AEFC) was developed for the measurement of spent fuel rods and assemblies from research reactors for safeguards verification. This measurement system contains components for active neutron interrogation, passive neutron totals counting, neutron coincidence counting, and gross gamma-ray counting. This report presents the first application of the time correlated interrogation technique for the measurement of the 235 U content in research reactor spent fuel assemblies. The technique, called time correlated induced fission (TCIF), uses a 252 Cf neutron source to irradiate the fuel assembly, and the subsequent induced fission events in the fissile material are measured by coincidence counting. The doubles rates are enhanced by having the neutron trigger events from both the 252 Cf source and the induced fission neutrons in the same time gate in the coincidence analysis. The average neutrons per fission of the 252 Cf source is 3.76 and the induced fission neutrons for 235 U is 2.44, so the number of neutrons that are produced is higher than for random neutron interrogation. This high effective neutron number increases the multiplicity counting rates and reduces the statistical error. The background coincidence counts from the 252 Cf are reduced by the water in the sample cavity and the polyethylene surrounding the 3 He detector tubes. This method of active neutron interrogation has been applied to the measurement of spent research reactor (IRT) fuel assemblies. The advanced experimental fuel counter (AEFC) was used to compare the TCIF method with the typically used AmLi neutron interrogation source that emits neutrons that are random in time. The statistical uncertainty for the use of the random neutron source (AmLi) and the time correlated source ( 252 Cf) for spent fuel interrogations was evaluated.

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗
NASA NTRS1982

H0, q0 and the local velocity field

An attempt is made to find a systematic deviation from linearity for distances that are under the control of the Virgo cluster, and to determine the value of the mean random motion about the systematic flow, in order to improve the measurement of the Hubble and the deceleration constants. The velocity-distance relation for large and intermediate distances is studied, and type I supernovae are calibrated relatively as distance indicators and absolutely to obtain a new value for the Hubble constant. Methods of determining the deceleration constant are assessed, including determination from direct measurement, mean luminosity density, virgocentric motion, and the time scale test. The very local velocity field is investigated, and a solution is preferred with a random peculiar radial velocity of very nearby field galaxies of 90-100 km/s, and a Virgocentric motion of the local group of 220 km/s, leading to an underlying expansion rate of 55, in satisfactory agreement with the global value.

Sandage, A.↗
NASA NTRSDate not supplied

Combinatorial Reasoning: Selecting Reasons in Generative AI Pipelines via Combinatorial Optimization

Recent Large Language Models (LLMs) have demonstrated impressive capabilities at tasks that require human intelligence and are a significant step towards human-like artificial intelligence (AI). Yet the performance of LLMs at reasoning tasks have been subpar and the reasoning capability of LLMs is a matter of significant debate. While it has been shown that the choice of the prompting technique to the LLM can alter its performance on a multitude of tasks, including reasoning, the best performing techniques require human-made prompts with the knowledge of the tasks at hand. We introduce a framework for what we call Combinatorial Reasoning (CR), a fully-automated prompting method, where reasons are sampled from an LLM pipeline and mapped into a Quadratic Unconstrained Binary Optimization (QUBO) problem. The framework investigates whether QUBO solutions can be profitably used to select a useful subset of the reasons to construct a Chain-of-Thought style prompt. We explore the acceleration of CR with specialized solvers. We also investigate the performance of simpler zero-shot strategies such as linear majority rule or random selection of reasons. Our preliminary study indicates that coupling a combinatorial solver to generative AI pipelines is an interesting avenue for AI reasoning and elucidates design principles for future CR methods.

combinatorial reasoning↗
NASA NTRS2016

Optics of Water Microdroplets with Soot Inclusions: Exact Versus Approximate Results

We use the recently generalized version of the multi-sphere superposition T-matrix method (STMM) to compute the scattering and absorption properties of microscopic water droplets contaminated by black carbon. The soot material is assumed to be randomly distributed throughout the droplet interior in the form of numerous small spherical inclusions. Our numerically-exact STMM results are compared with approximate ones obtained using the Maxwell-Garnett effective-medium approximation (MGA) and the Monte Carlo ray-tracing approximation (MCRTA). We show that the popular MGA can be used to calculate the droplet optical cross sections, single-scattering albedo, and asymmetry parameter provided that the soot inclusions are quasi-uniformly distributed throughout the droplet interior, but can fail in computations of the elements of the scattering matrix depending on the volume fraction of soot inclusions. The integral radiative characteristics computed with the MCRTA can deviate more significantly from their exact STMM counterparts, while accurate MCRTA computations of the phase function require droplet size parameters substantially exceeding 60.

cloud droplets↗
DOE OSTI2023

Data-driven analysis and prediction of stable phases for high-entropy alloy design

High-entropy alloys (HEAs) represent a promising class of materials with exceptional structural and functional properties. However, their design and optimization pose challenges due to the large composition-phase space coupled with the complex and diverse nature of the phase formation dynamics. In this study, a data-driven approach that utilizes machine learning (ML) techniques to predict HEA phases and their composition-dependent phases is proposed. By employing a comprehensive dataset comprising 5692 experimental records encompassing 50 elements and 11 phase categories, we compare the performance of various ML models. Our analysis identifies the most influential features for accurate phase prediction. Furthermore, the class imbalance is addressed by employing data augmentation methods, raising the number of records to 1500 in each category, and ensuring a balanced representation of phase categories. The results show that XGBoost and Random Forest consistently outperform the other models, achieving 86% accuracy in predicting all phases. Additionally, this work provides an extensive analysis of HEA phase formers, showing the contributions of elements and features to the presence of specific phases. We also examine the impact of including different phases on ML model accuracy and feature significance. Notably, the findings underscore the need for ML model selection based on specific applications and desired predictions, as feature importance varies across models and phases. This study significantly advances the understanding of HEA phase formation, enabling targeted alloy design and fostering progress in the field of materials science.

36 MATERIALS SCIENCE↗
DOE OSTI2025

SymbolFit: Automatic Parametric Modeling with Symbolic Regression

We introduce SymbolFit (API: https://github.com/hftsoi/symbolfit), a framework that automates parametric modeling by using symbolic regression to perform a machine-search for functions that fit the data while simultaneously providing uncertainty estimates in a single run. Traditionally, constructing a parametric model to accurately describe binned data has been a manual and iterative process, requiring an adequate functional form to be determined before the fit can be performed. The main challenge arises when the appropriate functional forms cannot be derived from first principles, especially when there is no underlying true closed-form function for the distribution. In this work, we develop a framework that automates and streamlines the process by utilizing symbolic regression, a machine learning technique that explores a vast space of candidate functions without requiring a predefined functional form because the functional form itself is treated as a trainable parameter, making the process far more efficient and effortless than traditional regression methods. We demonstrate the framework in high-energy physics experiments at the CERN Large Hadron Collider (LHC) using five real proton-proton collision datasets from new physics searches, including background modeling in resonance searches for high-mass dijet, trijet, paired-dijet, diphoton, and dimuon events. We show that our framework can flexibly and efficiently generate a wide range of candidate functions that fit a nontrivial distribution well using a simple fit configuration that varies only by random seed, and that the same fit configuration, which defines a vast function space, can also be applied to distributions of different shapes, whereas achieving a comparable result with traditional methods would have required extensive manual effort.

Tsoi, Ho Fung [Univ. of Pennsylvania, Philadelphia↗
NASA NTRS1995

Searching for patterns in remote sensing image databases using neural networks

We have investigated a method, based on a successful neural network multispectral image classification system, of searching for single patterns in remote sensing databases. While defining the pattern to search for and the feature to be used for that search (spectral, spatial, temporal, etc.) is challenging, a more difficult task is selecting competing patterns to train against the desired pattern. Schemes for competing pattern selection, including random selection and human interpreted selection, are discussed in the context of an example detection of dense urban areas in Landsat Thematic Mapper imagery. When applying the search to multiple images, a simple normalization method can alleviate the problem of inconsistent image calibration. Another potential problem, that of highly compressed data, was found to have a minimal effect on the ability to detect the desired pattern. The neural network algorithm has been implemented using the PVM (Parallel Virtual Machine) library and nearly-optimal speedups have been obtained that help alleviate the long process of searching through imagery.

Paola, Justin D.↗
NASA NTRS2011

Stress-Driven Selection of Novel Phenotypes

A process has been developed that can confer novel properties, such as metal resistance, to a host bacterium. This same process can also be used to produce RNAs and peptides that have novel properties, such as the ability to bind particular compounds. It is inherent in the method that the peptide or RNA will behave as expected in the target organism. Plasmid-born mini-gene libraries coding for either a population of combinatorial peptides or stable, artificial RNAs carrying random inserts are produced. These libraries, which have no bias towards any biological function, are used to transform the organism of interest and to serve as an initial source of genetic variation for stress-driven evolution. The transformed bacteria are propagated under selective pressure in order to obtain variants with the desired properties. The process is highly distinct from in vitro methods because the variants are selected in the context of the cell while it is experiencing stress. Hence, the selected peptide or RNA will, by definition, work as expected in the target cell as the cell adapts to its presence during the selection process. Once the novel gene, which produces the sought phenotype, is obtained, it can be transferred to the main genome to increase the genetic stability in the organism. Alternatively, the cell line can be used to produce novel RNAs or peptides with selectable properties in large quantity for separate purposes. The system allows for easy, large-scale purification of the RNAs or peptide products. The process has been reduced to practice by imposing sub-inhibitory concentrations of NiCl2 on cells of the bacterium Escherichia coli that were transformed separately with the peptide library and RNA library. The evolved resistant clones were isolated, and sequences of the selected mini-gene variants were established. Clones resistant to NiCl2 were found to carry identical plasmid variants with a functional mini-gene that specifically conferred significant nickel tolerance on the host cells. Sequencing of the selected mini-gene revealed a propensity of the encoded peptide to bind transient metal ions. Expression of the mini-gene markedly improved growth parameters of the evolved clones at sub-inhibitory concentrations of NiCl2 while being slightly detrimental in the absence of stress. Similar results have been obtained with the RNA libraries. Overall, the results demonstrate a very natural outcome of the selection experiments in which the mini-genes were expected to be either successfully integrated into bacterial genetic networks, or rejected depending upon their effect on host fitness. This described approach can be useful as a laboratory model to study the dynamics of bacterial adaptive evolution on the molecular level. It can also provide a strategy for screening expressed DNA libraries in search of novel genes with desirable properties.

Fox, George E.↗
NASA NTRS1980

Performability evaluation of the SIFT computer

The paper deals with the models, techniques, and evaluation methods that were successfully used to test the performance of the SIFT degradable computing system. The performance of the computer plus its air transport mission environment is modeled as a random variable, taking values in a set of 'accomplishment level'. The levels are defined in terms of four attributes of total system (computer plus environment) behavior, namely safety, no change in mission profile, no operational penalties, and no economic penalties. The base model of the total system is a stochastic process, whose states describe the internal structure of SIFT and the relevant conditions of its computational environment. Base model state trajectories are related to accomplishment levels via a special function, and solution methods are then used to determine the performability of the total system for various parameters of the computer and environment.

Meyer, J. F.↗
DOE OSTI2023

Quantum circuit fidelity estimation using machine learning

The computational power of real-world quantum computers is limited by errors. When using quantum computers to perform algorithms which cannot be efficiently simulated classically, it is important to quantify the accuracy with which the computation has been performed. In this work, we introduce a machine learning-based technique to estimate the fidelity between the state produced by a noisy quantum circuit and the target state corresponding to ideal noise-free computation. Our machine learning model is trained in a supervised manner, using smaller or simpler circuits for which the fidelity can be estimated using other techniques like direct fidelity estimation and quantum state tomography. Here we demonstrate that, for simulated random quantum circuits with a realistic noise model, the trained model can predict the fidelities of more complicated circuits for which such methods are infeasible. In particular, we show that the trained model may make predictions for circuits with higher degrees of entanglement than were available in the training set and that the model may make predictions for non-Clifford circuits even when the training set included only Clifford-reducible circuits. This empirical demonstration suggests classical machine learning may be useful for making predictions about beyond-classical quantum circuits for some non-trivial problems.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Using electronic health record metadata to predict housing instability amongst veterans

Housing instability is considered a significant life stressor and preemptive screening should be applied to identify those at risk for homelessness as early as possible so that they can be targeted for specialized care. We developed models to classify patient outcomes for an established VA Homelessness Screening Clinical Reminder (HSCR), which identifies housing instability, in the two months prior to its administration. Logistic Regression and Random Forest models were fit to classify responses using the last 18 months of document activity. We measure concentration of risk across stratifications of predicted probability and observe an enriched likelihood of finding confirmed false negative responses from veterans with diagnosed housing instability. Positive responses were 34 times more likely to be detected within the top 1 % of patients predicted at risk than from those randomly selected. There is a 1 in 4 chance of detecting false negatives within the top 1 % of predicted risk. Machine learning methods can classify between episodes of housing instability using a data-driven approach that does not rely on variables curated from domain experts. This method has the potential to improve clinicians’ ability to identify veterans who are experiencing housing instability but are not captured by HSCR.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2020

Learning and meta-learning of stochastic advection–diffusion–reaction systems from sparse measurements

Physics-informed neural networks (PINNs) were recently proposed in [18] as an alternative way to solve partial differential equations (PDEs). A neural network (NN) represents the solution, while a PDE-induced NN is coupled to the solution NN, and all differential operators are treated using automatic differentiation. Here, we first employ the standard PINN and a stochastic version, sPINN, to solve forward and inverse problems governed by a non-linear advection–diffusion–reaction (ADR) equation, assuming we have some sparse measurements of the concentration field at random or pre-selected locations. Subsequently, we attempt to optimise the hyper-parameters of sPINN by using the Bayesian optimisation method (meta-learning) and compare the results with the empirically selected hyper-parameters of sPINN. In particular, for the first part in solving the inverse deterministic ADR, we assume that we only have a few high-fidelity measurements, whereas the rest of the data is of lower fidelity. Hence, the PINN is trained using a composite multi-fidelity network, first introduced in [12], that learns the correlations between the multi-fidelity data and predicts the unknown values of diffusivity, transport velocity and two reaction constants as well as the concentration field. For the stochastic ADR, we employ a Karhunen–Loève (KL) expansion to represent the stochastic diffusivity, and arbitrary polynomial chaos (aPC) to represent the stochastic solution. Correspondingly, we design multiple NNs to represent the mean of the solution and learn each aPC mode separately, whereas we employ a separate NN to represent the mean of diffusivity and another NN to learn all modes of the KL expansion. For the inverse problem, in addition to stochastic diffusivity and concentration fields, we also aim to obtain the (unknown) deterministic values of transport velocity and reaction constants. The available data correspond to 7spatial points for the diffusivity and 20 space–time points for the solution, both sampled 2000 times. We obtain good accuracy for the deterministic parameters of the order of 1–2% and excellent accuracy for the mean and variance of the stochastic fields, better than three digits of accuracy. In the second part, we consider the previous stochastic inverse problem, and we use Bayesian optimisation to find five hyper-parameters of sPINN, namely the width, depth and learning rate of two NNs for learning the modes. Here, we obtain much deeper and wider optimal NNs compared to the manual tuning, leading to even better accuracy, i.e., errors less than 1% for the deterministic values, and about an order of magnitude less for the stochastic fields.

97 MATHEMATICS AND COMPUTING↗