Deep Learning Predictions Research Articles

Identification and Structural Characterization of Twisted Atomically Thin Bilayer Materials by Deep Learning.

Two-dimensional materials are expected to play an important role in next-generation electronics and optoelectronic devices. Recently, twisted bilayer graphene and transition metal dichalcogenides have attracted significant attention due to their unique physical properties and potential applications. In this study, we describe the use of optical microscopy to collect the color space of chemical vapor deposition (CVD) of molybdenum disulfide (MoS2) and the application of a semantic segmentation convolutional neural network (CNN) to accurately and rapidly identify thicknesses of MoS2 flakes. A second CNN model is trained to provide precise predictions on the twist angle of CVD-grown bilayer flakes. This model harnessed a data set comprising over 10,000 synthetic images, encompassing geometries spanning from hexagonal to triangular shapes. Subsequent validation of the deep learning predictions on twist angles was executed through the second harmonic generation and Raman spectroscopy. Our results introduce a scalable methodology for automated inspection of twisted atomically thin CVD-grown bilayers.

Physical consistency and invariance in machine learning of turbulent signals

This paper concerns an investigation of the invariance and consistency of deep learning of turbulent pressure fluctuations. The long-short-memory model is employed to predict wall pressure fluctuations across physical regimes featuring turbulence, shock–boundary layer interaction, and separation. The model's sensitivity to the data inputs is examined using different input data sets. Training the deep learning model based on the raw signals from different flow regions leads to large inaccuracies. It is shown that the data must be appropriately pre-processed before training for the deep learning model predictions to become consistent. Removing the mean and using the normalized fluctuating component of the signal, the deep learning predictions not only greatly improved in accuracy but, most importantly, converged and became consistent, provided that the signal sparsity remains within the inertial sub-range of the turbulence energy spectrum cascade. The power spectra of the surface pressure fluctuations reveal that the model provides high accuracy up to a certain frequency for the fully turbulent flow. The deep learning model's consistency is evidenced by being transferable across the various probe positions on the wall despite the significant differences in the turbulent flow properties in the training data set, i.e., signals obtained before, after, and inside the shock–boundary layer interaction regions. The model's prediction consistency and invariance to the turbulent signal training location(s) are promising for applying deep learning models to various turbulent flows.

Deep learning approaches for modeling laser-driven proton beams via phase-stable acceleration

Deep learning (DL) has recently become a powerful tool for optimizing parameters and predicting phenomena to boost laser-driven ion acceleration. We developed a neural network surrogate model using an ensemble of 355 one-dimensional particle-in-cell simulations to validate the theory of phase-stable acceleration (PSA) driven by a circularly polarized laser driver. Our DL predictions confirm the PSA theory and reveal a discrepancy in the required target density for stable ion acceleration at larger target thicknesses. We discuss the physical reasons behind this density underestimation based on our DL insights.

Deep learning for automated left ventricular outflow tract diameter measurements in 2D echocardiography

BackgroundMeasurement of the left ventricular outflow tract diameter (LVOTd) in echocardiography is a common source of error when used to calculate the stroke volume. The aim of this study is to assess whether a deep learning (DL) model, trained on a clinical echocardiographic dataset, can perform automatic LVOTd measurements on par with expert cardiologists.MethodsData consisted of 649 consecutive transthoracic echocardiographic examinations of patients with coronary artery disease admitted to a university hospital. 1304 LVOTd measurements in the parasternal long axis (PLAX) and zoomed parasternal long axis views (ZPLAX) were collected, with each patient having 1–6 measurements per examination. Data quality control was performed by an expert cardiologist, and spatial geometry data was preserved for each LVOTd measurement to convert DL predictions into metric units. A convolutional neural network based on the U-Net was used as the DL model.ResultsThe mean absolute LVOTd error was 1.04 (95% confidence interval [CI] 0.90–1.19) mm for DL predictions on the test set. The mean relative LVOTd errors across all data subgroups ranged from 3.8 to 5.1% for the test set. Generally, the DL model had superior performance on the ZPLAX view compared to the PLAX view. DL model precision for patients with repeated LVOTd measurements had a mean coefficient of variation of 2.2 (95% CI 1.6–2.7) %, which was comparable to the clinicians for the test set.ConclusionDL for automatic LVOTd measurements in PLAX and ZPLAX views is feasible when trained on a limited clinical dataset. While the DL predicted LVOTd measurements were within the expected range of clinical inter-observer variability, the robustness of the DL model requires validation on independent datasets. Future experiments using temporal information and anatomical constraints could improve valvular identification and reduce outliers, which are challenges that must be addressed before clinical utilization.Graphical

Equivalent Dose Estimation in FLASH Irradiation with a Deep Learning Approach

Ultra-high dose rate (FLASH) irradiation has been reported to provide decreased normal tissue toxicity without compromising tumor control compared with conventional (CONV) irradiation. However, a comprehensive understanding of the FLASH biological effect requires precise quantification of radiobiology. The study is to explore whether deep learning (DL) can tackle the task. As a proof of concept, we investigate a DL model for estimating FLASH dose to its equivalent CONV dose. Healthy C57Bl/6 female mice underwent FLASH (200Gy/s; n = 43) or CONV (0.12Gy/s; n = 41) whole abdominal irradiation using ∼16 MeV electron beams with a dose escalation scheme of 5 groups (n = 8 or 9) at 1Gy increments: 12-16Gy FLASH, 11-15Gy CONV. 4 days post-irradiation, 9 jejunum cross-sections per mouse were H&E stained for histological analysis. Each cross-section image was processed to remove lumen background and oversampled into multiple large-scale and small-scale patches along jejunal circumference. In CONV dataset, we randomly selected the data of 32 mice (80%) for model training and the rest (20%) for model validation. A ResNet101-based DL model, pre-trained with an unsupervised contrastive learning scheme, was retrained with only CONV training set to estimate corresponding CONV dose. For comparison, a crypt counting (CC) approach was implemented by manually counting the number of regenerating crypts on each cross-section image. An exponential function of dose vs crypt number was fitted with the CONV training set and used for dose estimation on the testing set. Mean squared error (MSE) was used to assess the accuracy of DL and CC approaches in estimating dose levels in CONV irradiation. The validated DL model was applied to the FLASH set to project FLASH dose into corresponding CONV dose that results in equivalent biological response. The CONV dose estimated by DL and CC approaches and DL-estimated FLASH equivalent dose were summarized in Table 1. The DL model achieved an MSE of 0.21 Gy2 on CONV testing set compared with 0.32 Gy2 of the CC approach. FLASH equivalent dose estimated by DL model for 12, 13, 14, 15 and 16Gy were 12.16±0.40, 12.53±0.32, 12.72±0.24, 12.85±0.20 and 13.04±0.27 Sv, respectively. Our proposed DL model can accurately estimate the CONV dose based on histological images. The DL predictions of FLASH dataset demonstrate that FLASH may reduce normal tissue toxicity with a lower equivalent dose, especially at high irradiated dose levels. Our study indicates that deep learning can be potentially used to assess the equivalent dose of FLASH irradiation to normal tissue.

Weeks-Ahead Epidemiological Predictions of Varicella Cases From Univariate Time Series Data Applying Artificial Intelligence

Abstract Background “Chickenpox” is a highly infectious disease caused by the varicella-zoster virus, influenced by seasonal and spatial factors. Dealing with varicella-zoster epidemics can be a substantial drain on health-authority resources. Methods that improve the ability to locally predict case numbers from time-series data sets every week are therefore worth developing. Methods Simple-to-extract trend attributes from published univariate weekly case-number univariate data sets were used to generate multivariate data for Hungary covering 10 years. That attribute-enhanced data set was assessed by machine learning (ML) and deep learning (DL) models to generate weekly case forecasts from next week (t0) to 12 weeks forward (t+12). The ML and DL predictions were compared with those generated by multilinear regression and univariate prediction methods. Results Support vector regression generates the best predictions for weeks t0 and t+1, whereas extreme gradient boosting generates the best predictions for weeks t+3 to t+12. Long-short-term memory only provides comparable prediction accuracy to the ML models for week t+12. Multi–K-fold cross validation reveals that overall the lowest prediction uncertainty is associated with the tree-ensemble ML models. Conclusion The novel trend-attribute method offers the potential to reduce prediction errors and improve transparency for chickenpox time series.

Approximating the uncertainty of deep learning reconstruction predictions in single-pixel imaging

Single-pixel imaging (SPI) has the advantages of high-speed acquisition over a broad wavelength range and system compactness. Deep learning (DL) is a powerful tool that can achieve higher image quality than conventional reconstruction approaches. Here, we propose a Bayesian convolutional neural network (BCNN) to approximate the uncertainty of the DL predictions in SPI. Each pixel in the predicted image represents a probability distribution rather than an image intensity value, indicating the uncertainty of the prediction. We show that the BCNN uncertainty predictions are correlated to the reconstruction errors. When the BCNN is trained and used in practical applications where the ground truths are unknown, the level of the predicted uncertainty can help to determine whether system, data, or network adjustments are needed. Overall, the proposed BCNN can provide a reliable tool to indicate the confidence levels of DL predictions as well as the quality of the model and dataset for many applications of SPI.

Deep learning reconstruction of pressure fluctuations in supersonic shock–boundary layer interaction

The long short-term memory deep-learning model is applied to supersonic shock–boundary layer interaction flow. The study aims to show how near-wall pressure fluctuations can be reconstructed from reduced (under-sampled) datasets of pressure signals. Predicting pressure fluctuations from reduced datasets could allow predictions using less expensive simulations and experiments. The training of the deep learning model is based on direct numerical simulations of supersonic ramp flows, focusing on the regions upstream of and around the shock–boundary layer interaction region. During the pre-processing stage, cubic spline functions increase the fidelity of the sparse signals and feed them to the long-short memory model for an accurate reconstruction. Comparisons are also carried out for different sparsity factors and assess the model's accuracy both qualitatively through the pressure signals and quantitatively using the root mean square error and the power spectra. The deep learning predictions are promising and can be extended to include other aerodynamic or aeroelastic parameters of interest.

Prediction ofsecondary electron yield formetal materials using deep learning.

This article describes a neural network system for predicting the secondary electron yield of metallic materials. For bulk metals, experimental values are used as training data. Due to the strong correlation between the secondary electron yield and the work function, deep learning predicts the secondary electron yield with relatively high accuracy even with a small amount of training data. Our approach demonstrates the importance of the work function in predicting the secondary electron yield. For the secondary electron yield of thin metal films on metal substrates, deep learning predictions are generated using training data obtained by Monte Carlo simulations. The accuracy of the secondary yield predictions of thin films on substrates could be improved by adding experimental values of bulk metals to the training data.

Dynamic Assimilation of Deep Learning Predictions to a Process-Based Water Budget

A three-step data assimilation (DA) of deep learning (DL) predictions to a process-based water budget is developed and applied to produce an active, operational water balance for groundwater management. In the first step, an existing water budget model provides forward model predictions of aquifer storage from meteorological observations, estimates of pumping and diversion discharge, and estimates of recharge. A Kalman filter DA approach is the second step and generates updated storage volumes by combining a long short-term memory (LSTM) network, a DL method, and predicted “measurements” with forward model predictions. The third “correction” step uses modified recharge and pumping, adjusted to account for the difference between Kalman update storage and forward model predicted storage, in forward model re-simulation to approximate updated storage volume. Use of modified inputs in the correction provides a mass-conservative water budget framework that leverages DL predictions. LSTM predictor “measurements” primarily represent missing observations due to missing or malfunctioning equipment. Pumping and recharge inputs are uncertain and unobserved in the study region and can be adjusted without contradicting measurements. Because DL requires clean and certain data for learning, a common-sense baseline facilitates interpretation of LSTM generalization skill and accounts for feature and outcome uncertainty when sufficient target data are available. DA, in contrast to DL, provides for explicit uncertainty analysis through an observation error model, which allows the integrated approach to address uncertainty impacts from an LSTM predictor developed from limited outcome observations.

Persistent topological Laplacian analysis of SARS-CoV-2 variants.

Topological data analysis (TDA) is an emerging field in mathematics and data science. Its central technique, persistent homology, has had tremendous success in many science and engineering disciplines. However, persistent homology has limitations, including its inability to handle heterogeneous information, such as multiple types of geometric objects; being qualitative rather than quantitative, e.g., counting a 5-member ring the same as a 6-member ring, and a failure to describe non-topological changes, such as homotopic changes in protein-protein binding. Persistent topological Laplacians (PTLs), such as persistent Laplacian and persistent sheaf Laplacian, were proposed to overcome the limitations of persistent homology. In this work, we examine the modeling and analysis power of PTLs in the study of the protein structures of the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) spike receptor binding domain (RBD). First, we employ PTLs to study how the RBD mutation-induced structural changes of RBD-angiotensin-converting enzyme 2 (ACE2) binding complexes are captured in the changes of spectra of the PTLs among SARS-CoV-2 variants. Additionally, we use PTLs to analyze the binding of RBD and ACE2-induced structural changes of various SARS-CoV-2 variants. Finally, we explore the impacts of computationally generated RBD structures on a topological deep learning paradigm and predictions of deep mutational scanning datasets for the SARS-CoV-2 Omicron BA.2 variant. Our results indicate that PTLs have advantages over persistent homology in analyzing protein structural changes and provide a powerful new TDA tool for data science.

Time-dependent deep learning predictions of 3D electrode particle-resolved microstructure effect on voltage discharge curves

Battery electrode particle-resolved microstructure featured with a 3D active material (AM) particle network bonded by carbon and binder (CBD) phase plays a significant role in determining battery performances. Tremendous physical computational models were developed to investigate these microstructure effects on voltage discharge curves, which however are computationally expensive. Machine learning (ML) could be the solution to address the high computational cost, but they still suffer from two challenges. (1) 3D particle-resolved microstructure that cannot be modelled in traditional ML that only correlate the scalar inputs and outs, (2) the outputs (voltage discharge curves), are time dependent. In this work, a deep learning (DL) approach was developed to tackle these two challenges. Using NMC 111 as one example, the results show a great match between the predictions and ground truth with a 99.98% validation accuracy. The superb extrapolation capabilities were demonstrated through the accurate prediction of microstructures fabricated with distinct process conditions and graded microstructures that exhibit completely different probability density function (PDF) of particle size. The trained DL model can be employed to boost the optimization of fabrication process parameters and microstructure design of electrodes. Moreover, the proposed framework can be extended to predict other dynamic behaviors of particle-resolved electrode microstructures.

Residue-level error detection in cryoelectron microscopy models

Building accurate protein models into moderate resolution (3-5Å) cryoelectron microscopy (cryo-EM) maps is challenging and error prone. We have developed MEDIC (Model Error Detection in Cryo-EM), a robust statistical model that identifies local backbone errors in protein structures built into cryo-EM maps by combining local fit-to-density with deep-learning-derived structural information. MEDIC is validated on a set of 28 structures that were subsequently solved to higher resolutions, where we identify the differences between low- and high-resolution structures with 68% precision and 60% recall. We additionally use this model to fix over 100 errors in 12 deposited structures and to identify errors in 4 refined AlphaFold predictions with 80% precision and 60% recall. As modelers more frequently use deep learning predictions as a starting point for refinement and rebuilding, MEDIC's ability to handle errors in structures derived from hand-building and machine learning methods makes it a powerful tool for structural biologists.

Seg2Link: an efficient and versatile solution for semi-automatic cell segmentation in 3D image stacks

Recent advances in microscopy techniques, especially in electron microscopy, are transforming biomedical studies by acquiring large quantities of high-precision 3D cell image stacks. To examine cell morphology and connectivity in organs such as the brain, scientists need to conduct cell segmentation, which extracts individual cell regions of different shapes and sizes from a 3D image. This is challenging due to the indistinct images often encountered in real biomedical research: in many cases, automatic segmentation methods inevitably contain numerous mistakes in the segmentation results, even when using advanced deep learning methods. To analyze 3D cell images effectively, a semi-automated software solution is needed that combines powerful deep learning techniques with the ability to perform post-processing, generate accurate segmentations, and incorporate manual corrections. To address this gap, we developed Seg2Link, which takes deep learning predictions as inputs and use watershed 2D + cross-slice linking to generate more accurate automatic segmentations than previous methods. Additionally, it provides various manual correction tools essential for correcting mistakes in 3D segmentation results. Moreover, our software has been optimized for efficiently processing large 3D images in diverse organisms. Thus, Seg2Link offers an practical solution for scientists to study cell morphology and connectivity in 3D image stacks.

Generalising uncertainty improves accuracy and safety of deep learning analytics applied to oncology

Uncertainty estimation is crucial for understanding the reliability of deep learning (DL) predictions, and critical for deploying DL in the clinic. Differences between training and production datasets can lead to incorrect predictions with underestimated uncertainty. To investigate this pitfall, we benchmarked one pointwise and three approximate Bayesian DL models for predicting cancer of unknown primary, using three RNA-seq datasets with 10,968 samples across 57 cancer types. Our results highlight that simple and scalable Bayesian DL significantly improves the generalisation of uncertainty estimation. Moreover, we designed a prototypical metric—the area between development and production curve (ADP), which evaluates the accuracy loss when deploying models from development to production. Using ADP, we demonstrate that Bayesian DL improves accuracy under data distributional shifts when utilising ‘uncertainty thresholding’. In summary, Bayesian DL is a promising approach for generalising uncertainty, improving performance, transparency, and safety of DL models for deployment in the real world.

Deep learning predictions of unsteady aerodynamic loads on an airfoil model pitched over the entire operating range

For the design and certification of wind turbines, it is essential to provide fast and accurate unsteady aerodynamic load prediction models for the whole operational range of angle of attack, up to 180° for vertical-axis and 90° for horizontal-axis wind turbines. This work describes a computationally efficient unsteady forces prediction model based on a deep learning approach, namely the bidirectional long short-term memory (BiLSTM) algorithm, for an airfoil pitched over the full operational range of angles of attack up to 180°. No model has been developed to capture the unsteady forces at high angles of attack. Novel features based on operating conditions and the steady polars of the airfoil are used as inputs for the BiLSTM model. Direct measurements of steady and unsteady forces on a NACA 0021 airfoil model were conducted at reduced frequencies up to 0.075 and a Reynolds number of 120 000 in an open-jet wind tunnel for model learning and testing. The unsteady forces vary significantly from the steady values at high pitching amplitudes and post-stall angles, which, if not accounted for when simulating wind turbine performance, would result in inaccurate predictions. Furthermore, measurements revealed the effect of unsteady vorticity development and shedding on aerodynamic forces under forward and reverse flow conditions. The BiLSTM model is capable of capturing the underlying physics of unsteady aerodynamic forces under extreme operating conditions.

Computational approaches streamlining drug discovery.

Computer-aided drug discovery has been around for decades, although the past few years have seen a tectonic shift towards embracing computational technologies in both academia and pharma. This shift is largely defined by the flood of data on ligand properties and binding to therapeutic targets and their 3D structures, abundant computing capacities and the advent of on-demand virtual libraries of drug-like small molecules in their billions. Taking full advantage of these resources requires fast computational methods for effective ligand screening. This includes structure-based virtual screening of gigascale chemical spaces, further facilitated by fast iterative screening approaches. Highly synergistic are developments in deep learning predictions of ligand properties and target activities in lieu of receptor structure. Here we review recent advances in ligand discovery technologies, their potential for reshaping the whole process of drug discovery and development, as well as the challenges they encounter. We also discuss how the rapid identification of highly diverse, potent, target-selective and drug-like ligands to protein targets can democratize the drug discovery process, presenting new opportunities for the cost-effective development of safer and more effective small-molecule treatments.

A deep learning-based approach for statistical robustness evaluation in proton therapy treatment planning: a feasibility study

Objective. Robustness evaluation is critical in particle radiotherapy due to its susceptibility to uncertainties. However, the customary method for robustness evaluation only considers a few uncertainty scenarios, which are insufficient to provide a consistent statistical interpretation. We propose an artificial intelligence–based approach that overcomes this limitation by predicting a set of percentile dose values at every voxel and allows for the evaluation of planning objectives at specific confidence levels. Approach. We built and trained a deep learning (DL) model to predict the 5th and 95th percentile dose distributions, which corresponds to the lower and upper bounds of a two-tailed 90% confidence interval (CI), respectively. Predictions were made directly from the nominal dose distribution and planning computed tomography scan. The data used to train and test the model consisted of proton plans from 543 prostate cancer patients. The ground truth percentile values were estimated for each patient using 600 dose recalculations representing randomly sampled uncertainty scenarios. For comparison, we also tested whether a common worst-case scenario (WCS) robustness evaluation (voxel-wise minimum and maximum) corresponding to a 90% CI could reproduce the ground truth 5th and 95th percentile doses. Main results. The percentile dose distributions predicted by DL yielded excellent agreements with the ground truth dose distributions, with mean dose errors below 0.15 Gy and average gamma passing rates (GPR) at 1 mm/1% above 93.9, which were substantially better than the WCS dose distributions (mean dose error above 2.2 Gy and GPR at 1 mm1% below 54). We observed similar outcomes in a dose-volume histogram error analysis, where the DL predictions generally yielded smaller mean errors and standard deviations than the WCS evaluation doses. Significance. The proposed method produces accurate and fast predictions (∼2.5 s for one percentile dose distribution) for a given confidence level. Thus, the method has the potential to improve robustness evaluation.

Data-driven methods for stress field predictions in random heterogeneous materials

Predicting full-field stress responses is of fundamental importance to assessing materials failure and has various engineering applications in design optimization, manufacturing process control, and structural health monitoring. This article develops and evaluates different data-driven methods for efficient and accurate predictions of full stress fields in random heterogeneous materials. The first approach integrates model order reduction of proper orthogonal decomposition (POD) with classical machine learning techniques (K-nearest neighbors, random forest, and artificial neural networks) to predict full-field responses based on POD-reduced coefficients. However, this strategy shows limitations in predicting full stress fields, especially for heterogeneous material inclusions of small size or being close to the domain boundary. After that, two computer vision-based deep learning approaches were developed for full-field predictions. The first one uses a Resnet-based Convolutional Neural Network (CNN), and the second is based on a modified conditional Generative Adversarial Network (cGAN). Two representative example problems were studied: a random heterogeneous material inclusion or a void varying in size and location. In contrast to POD-based classical machine learning, almost invisible differences were found between the entire stress fields in finite element simulations and computer vision-based deep learning (CNN/cGAN) predictions, with significantly reduced mean squared error (MSE) and correlation values (R2) mostly above 0.99. On the other hand, the proposed cGAN provides more accurate predictions than CNN with fewer epochs.

Controlling auxeticity in curved-beam metamaterials via a deep generative model

Lattice-based mechanical metamaterials are known to exhibit quite a unique mechanical behavior owing to their rational internal architecture. This includes unusual properties such as a negative Poisson’s ratio, which can be easily tuned in reentrant-hexagonal metamaterials by adjusting the angles between beams. However, changing the angles also affects the geometrical dimensions of the unit cell. We show that by replacing traditional straight beams with curved ones, it is possible to control Poisson’s ratio of reentrant-hexagonal metamaterials keeping their overall dimensions constant. While the mechanical properties of these structures can be predicted through finite element simulations or, in some cases, analytically, many applications require the identification of architectures with specific target properties. To solve this inverse problem, we introduce a deep learning framework for generating metamaterials with desired properties. By supplying the generative model with a guide structure in addition to the target properties, we are able to generate a large number of alternative architectures with the same properties and express a preference for a specific shape. Deep learning predictions, together with experimental measurements, prove that this approach allows us to accurately generate unit cells fitting specific properties for curved-beam metamaterials.