Fast Prediction Research Articles

BrainQCNet: A Deep Learning attention-based model for the automated detection of artifacts in brain structural MRI scans

Abstract Analyses of structural MRI (sMRI) data depend on robust upstream data quality control (QC). It is also crucial that researchers seek to retain maximal amounts of data to ensure reproducible, generalizable models and to avoid wasted effort, including that of participants. The time-consuming and difficult task of manual QC evaluation has prompted the development of tools for the automatic assessment of brain sMRI scans. Existing tools have proved particularly valuable in this age of Big Data; as datasets continue to grow, reducing execution time for QC evaluation will be of considerable benefit. The development of Deep Learning (DL) models for artifact detection in structural MRI scans offers a promising avenue toward fast, accurate QC evaluation. In this study, we trained an interpretable Deep Learning model, ProtoPNet, to classify minimally preprocessed 2D slices of scans that had been manually annotated with a refined quality assessment (ABIDE 1; n = 980 scans). To evaluate the best model, we applied it to 2141 ABCD T1-weighted MRI scans for which gold-standard manual QC annotations were available. We obtained excellent accuracy: 82.4% for good quality scans (Pass), 91.4% for medium to low quality scans (Fail). Further validation using 799 T1w MRI scans from ABIDE 2 and 750 T1w MRI scans from ADHD-200 confirmed the reliability of our model. Accuracy was comparable to or exceeded that of existing ML models, with fast processing and prediction time (1 minute per scan, GPU machine, CUDA-compatible). Our attention model also performs better than traditional DL (i.e., convolutional neural network models) in detecting poor quality scans. To facilitate faster and more accurate QC prediction for the neuroimaging community, we have shared the model that returned the most reliable global quality scores as a BIDS-app (https://github.com/garciaml/BrainQCNet).

Evaluation of supervised machine learning regression models for CFD-based surrogate modelling in indoor airflow field reconstruction

Fast and reliable prediction of indoor airflow distribution is critical for indoor environment control. While neural networks (NN), often interchangeably referred to as Back Propagation Neural Networks (BPNNs), are popular for airflow predictions, optimizing these models is challenging due to their ”black box” nature and complex network structures. This study explores alternative robust regression models, including decision-tree-based models (e.g., XGBoost, LightGBM, Random Forest) and Support Vector Regression (SVR), for predicting indoor airflow. Two BPNN structures were initially developed to evaluate feasibility of NN models. BPNN A was trained using airflow velocities from two inlets as input neurons to directly predict the airflow velocity distribution within the domain. BPNN B was trained additionally with spatial information, including space samples and boundary wall data. Higher-dimensional training structures of BPNN B were applied to decision tree-based models and SVR to assess their capability in predicting non-linear airflow patterns. Results indicated that BPNN A achieved the highest accuracy, while the inclusion of higher-dimensional data in BPNN B led to decreased accuracy. Among all decision-tree-based models, XGBoost demonstrated the greatest potential, achieving an R2 above 99.5% and predictive errors below 10%. XGBoost also outperformed both BPNN models in speed, being 15.78 times faster than BPNN A and 252 times faster than BPNN B. The interpretability of XGBoost was further explored by analysing feature importance, which helps identify the most influential input variables while predicting the airflow velocity. This analysis is expected to offer an enhanced understanding of boundary conditions leading to optimised indoor environment strategy.

Risk assessment of gas plume dispersion in bunkering of alternative fuel through validated computational fluid dynamics approach

Abstract Maritime decarbonization is crucial in fighting climate change. Adopting low-carbon alternative fuels could help reduce greenhouse gas (GHG) emissions by 50% by 2050 compared to 2008. While liquified natural gas (LNG) has gained a strong foothold in the maritime sector over the past decades, the trend is shifting toward adopting lower and zero-carbon hydrogen and its carriers, such as methanol and ammonia, as alternative fuels. Quantifying risks from accidental leakages during storage and transfer of these alternative fuels using numerical models is required and commonly adopted in industries. The process leverages multiple models and tools for risk assessments with different accuracy levels and costs ranging from integral models to Lagrangian techniques and Eulerian approaches such as computational fluid dynamics (CFD). Increasingly, CFD approaches are employed for near-field dispersion analysis to assess the risk of accidental leakages. Despite their high computational cost, CFD models offer better accuracy than integral and Lagrangian models. In this work, we developed a CFD-based framework as a quantitative tool for risk assessment purposes. Our framework aims to (1) offer finer gas dispersion resolution with a detailed representation of surrounding structures and geometries, (2) minimize reliance on assumptions and empirical models to capture flow dynamics and gas dispersions accurately, and (3) mitigate conservative estimates and over-design of safety measures. The framework was validated using experimental data from various field tests, including a recent one involving releasing cryogenic liquids into the atmosphere. After successfully validating and comparing with experiments and other available tools, the proposed framework investigated gas plume dispersion from accidental leakages during fuel bunkering operations in ship-to-ship and shore-to-ship scenarios. The data obtained from CFD simulations was subsequently utilized to construct a surrogate model for fast prediction of gas plume behavior under varying environmental conditions. This facilitated the development of an effective emergency response plan.

Multi-dimensional distribution prediction of PM2.5 concentration in urban residential areas based on CNN

Air pollution has been one of the major environmental health issues in urban residential areas. Enhancing the air quality of the human environment has become a relevant research topic, but the study of the distribution characteristics of air pollutants under the influence of multi-dimensional indicators of residential building design is rarely considered. Herein, the PM2.5 spatial distribution of residential environment in Hefei, China, was simulated using ENVI-met. This research explored the influence mechanism of different design indexes on PM2.5 in the residential environment. A fast prediction method was proposed for PM2.5 based on the convolutional neural network (CNN). Furthermore, 4183 test points in the residential environment were selected to explore the influence of design indexes on PM2.5. The morphological features around the test points were transform to the images as input datasets for CNN model training. Results indicated that the PM2.5 mass concentration in the residential area decreases with the increase in height, and it is correlated with the sky viewfactor, the average height of the building and the density of the building. The accuracy of the CNN prediction model exceeds that of the ANN prediction model by 14.2 %, and the results can effectively predict the multi-dimensional distribution of the PM2.5 mass concentration in the residential environment, which could provide an effective reference and analytical tool for the creation of a healthy environment in residential areas.

Machine learning-based surrogate models for fast impact assessment of a new building on urban local microclimate at design stage

The rapid urbanization introduces changes in the local urban microclimate. Many efforts have been paid on the impact assessment of building design on local microclimate. However, there is still a lack of efficient and accurate prediction method for assessing the impacts on local microclimate when making the design of individual buildings. Two complementary machine learning-based surrogate models are proposed, including an SVR-based local air temperature model and a LightGBM-based local wind velocity model. They are identified by evaluating and comparing eight alternative machine learning models, four for each model development. 200 sets of CFD simulation data corresponding to different building designs are used for the model training and validation. The results show that the developed surrogate models can dramatically reduce computation time (from over 5 h to less than a second for a single prediction) while keeping the same order of accuracy of CFD simulations for local microclimate prediction of individual buildings. It therefore facilitates the fast, comprehensive and accurate prediction of the impacts on the local microclimate at the early design stage of new construction and renovation of individual buildings, for designers to deliver preferred local microclimate and/or avoid unacceptable microclimate changes.

TBM Advanced Geological Prediction via Ellipsoidal Positioning Velocity Analysis

Traditional seismic wave-based tunnel advanced geological forecasting techniques are primarily designed for drill and blast method construction tunnels. However, given the fast excavation speed and limited prediction space in tunnel boring machine (TBM) construction tunnels, traditional methods have significant technical limitations. This study analyzes the characteristics of different types of TBM construction tunnels and, considering the practical construction conditions, identifies an effective observation system and data acquisition method. To address the challenges in advanced forecasting for TBM construction tunnels, a method of ellipsoid positioning velocity analysis, which takes into account the constraints of three-component data directions, is proposed. Based on the characteristics of the advanced forecasting observation system, this method compares the maximum values on the spatial isochronous ellipsoidal surface to determine the average velocity of the geological layer rays, thereby enabling accurate inversion of the spatial distribution ahead. Utilizing numerical simulation, a model for the advanced detection of typical unfavorable geological formations is established by obtaining the wave field response characteristics of seismic waves in three-dimensional space, and the velocity structure of the model is retrieved through this velocity analysis method. In the engineering example, the fracture property, water content, and weathering degree of the surrounding rock are predicted accurately.

Coupling Dynamics Study on Multi-Body Separation Process of Underwater Vehicles

Based on the Newton-Euler method, a coupling rigid-body dynamics model of a Trans-Medium Vehicle (TMV) separating from an Unmanned Underwater Vehicle (UUV) has been established. The modeling is based on the “holistic method” and “Kane” ideas respectively, so that most of the equations can be derived without considering the internal forces between the two bodies. The separation propulsion force, which is an internal force, only appears in the relative glide dynamics equation of the TMV along the axis of the separation tube that is installed on the UUV. This greatly reduces the workload of modeling and derivation. The UUV works entirely underwater, while the hydrodynamic shape of the TMV changes continuously during the process of the TMV separating from the UUV. Therefore, accurate hydrodynamic calculations for the UUV and TMV are the basis of numerical resolution for the two rigid bodies’ coupling dynamics model in water. A large number of numerical simulations was conducted using CFD methods to investigate the hydrodynamic performance of the UUV and TMV under various conditions. These simulations aim to establish a hydrodynamic database, and accurate hydrodynamic models were developed through fitting methods and online interpolation. In the process of solving the coupling dynamics of two bodies, the hydrodynamic model is used to calculate the hydrodynamic force experienced by the UUV and TMV. This balances the accuracy and efficiency of a numerical simulation. Finally, numerous simulations and comparative analyses were conducted under various operational conditions and separation parameters. The simulation results indicate that the impact of TMV separation on the motion state of the UUV becomes more prominent with smaller UUV to TMV mass ratios or deeper TMV separation depths. This effect can further influence the stability control of the UUV. The coupling rigid body dynamics analysis method established in this paper provides a fast and effective prediction method for use during the scheme design and separation safety evaluation phases of creating UUV-TMV systems.

An optimal goose lithium-ion batteries accurate and rapid RUL prediction method with automatic initial hyperparameters settings

Abstract Machine learning has emerged as a highly effective tool for addressing complex data problems, garnering significant attention in the field of equipment degradation and remaining service life prediction. Existing prediction models typically exhibit two primary shortcomings: on the one hand, the accuracy of life prediction reaches the desired level of precision while failing to achieve a sufficiently fast prediction speed, and on the other hand, generalization is not guaranteed while requiring the model to be robust. These two aspects present a significant challenge to the field of machine learning. In light of the aforementioned issues, we propose a prediction model based on the goose algorithm. Initially, we set the goose algorithm using adaptive initialization of the goose population to guarantee that the goose population is set at the appropriate interval, and we incorporate it into the extreme learning machine model through the improved goose algorithm. goose algorithm is used to predict the service life. Finally, we utilize different types of lithium batteries with varying operational conditions to conduct pertinent case studies to validate the proposed prediction model. The results demonstrated that the average accuracy was above 98% in all validated datasets. The shortest computation time was 0.19 s.

Non-invasive, fast, and high-performance EGFR gene mutation prediction method based on deep transfer learning and model stacking for patients with Non-Small Cell Lung Cancer

PurposeTo propose an intelligent, non-invasive, highly precise, and rapid method to predict the mutation status of the Epidermal Growth Factor Receptor (EGFR) to accelerate treatment with Tyrosine Kinase Inhibitor (TKI) for patients with untreated adenocarcinoma Non-Small Cell Lung Cancer. Materials and methodsReal-world data from 521 patients with adenocarcinoma NSCLC who performed a CT scan and underwent surgery or pathological biopsy to determine EGFR gene mutation between January 2021 and July 2022, is collected. Solutions to the problems that prevent the model from achieving very high precision, namely: human errors made during the annotation of the database and the low precision of the output decision of the model, are proposed. Thus, among the 521 analyzed cases, only 40 were selected as patients with EGFR gene mutation and 98 cases with wild-type EGFR. ResultsThe proposed model is trained, validated, and tested on 12,040 2D images extracted from the 138 CT scans images where patients were randomly partitioned into training (80 %) and test (20 %) sets. The performance obtained for EGFR gene mutation prediction was 95.22 % for accuracy, 960.2 for F1_score, 95.89 % for precision, 96.92 % for sensitivity, 94.01 % for Cohen kappa, and 98 % for AUC. ConclusionAn EGFR gene mutation status prediction method, with high-performance thanks to an intelligent prediction model entrained by highly accurate annotated data is proposed. The outcome of this project will facilitate rapid decision-making when applying a TKI as an initial treatment.

A correlation-based approach for fast prediction of the equilibrium position of particles within a spiral microchannel

ABSTRACT As inertial microfluidics is a proven technique for manipulating particles within microchannels, a deep understanding of hydrodynamics forces acting on particles in these systems is crucial for optimal manipulation. However, existing numerical techniques for examining particle equilibrium positions are often time-intensive and require high-memory systems, limiting accessibility due to complexity and cost. To address this issue, this study proposes a method to develop a correlation for fast estimation of particle equilibrium positions. An algorithm is devised to develop the correlation in three main steps: first, a dataset is created by conducting simulations using the finite element method, considering width, height, flow rate, particle size, and curvature radius. The computed distance between the equilibrium position of the particle and the inner channel wall serves as input for the next step where a general form of the correlation is established using Buckingham’s Π-theorem. The unknown coefficients of the correlation are calculated using 80 percent of the dataset from the first step and the least sum squares errors method. Finally, the obtained correlation is validated using the remaining data and previous experimental studies. The proposed correlation reduces the computational burden associated with predicting particle focus positions, with a 3.7 percent mean absolute error.

Acousto-optic signal-based in-situ measurements supporting part quality improvement in additive manufacturing

In-situ measurements of process responses during powder bed fusion with laser beam of metals (PBF-LB/M) additive manufacturing are essential to monitor and control defect formation and improve product quality. Currently employed sensors are often unable to capture multiple responses together, affected by interferences, and insufficient to fully realize the transient nature of PBF-LB/M process. To address this issue, we developed a novel acousto-optic signal-based in-situ measurement system that can capture multiple signals. The interference in the signals due to the laser beam was uniquely analyzed and eliminated using statistical data analysis. The interference-free signals were found to provide insights into the formation of spatter and porosity. At high energy density, the spatter formation due to the liquid metal expulsion and powder ejection is the main factor affecting the porosity. A high energy density results in more splashing, and the porosity increases accordingly. The standard deviation of photovoltage signals and the energy distribution proportion estimated by wavelet decomposition showed a positive correlation with the spatter and porosity, respectively. The results showed that the processed signals can help in adjusting energy input to reduce spatter and porosity and improve part quality. In addition, a fast prediction of relative density based on the energy distribution characteristics of frequency zones was achieved using a back-propagation neural network with a correlation coefficient (R2) of 98.64%.

Tensile performance prediction of CFRPs with voids using multiscale analysis and neural networks

The overall work provides a novel multiscale modeling strategy applicable to the mechanical response prediction of arbitrarily laid porous laminates under uniaxial tension by establishing representative defective volumes (RDVs) for the inter-ply and intra-ply regions. A representative defective volume modeling method based on void characterization results using X-ray computed tomography for the interlaminar region was proposed. Elastic properties, tensile strength, and fracture energies of porous and poreless microscale models were obtained and used in the macroscale model. Defective properties were assigned to partial elements in the macroscale model according to the preset distribution mode. The Extended Finite Element Method was taken to simulate the failure process. Tensile properties of porous laminates with stacking sequences of [90]8, [90/0/90/0]2S, and [45/90/-45/0]2S were experimentally tested and numerically simulated. The predicted tensile properties and crack propagation behavior coincide well with the test results. Besides, the effects of micro-voids and macroscopic void distribution were discussed in detail. Further, a back-propagation neural network was used to achieve a fast prediction of the tensile properties and save computational costs. The computation time (<1 s) greatly improved compared with multiscale modeling (>38 h). This work provides an efficient tool for composite performance prediction, design, and optimization.

Screening of Silver-Based Single-Atom Alloy Catalysts for NO Electroreduction to NH3 by DFT Calculations and Machine Learning.

Exploring NO reduction reaction (NORR) electrocatalysts with high activity and selectivity toward NH3 is essential for both NO removal and NH3 synthesis. Due to their superior electrocatalytic activities, single-atom alloy (SAA) catalysts have attracted considerable attention. However, the exploration of SAAs is hindered by a lack of fast yet reliable prediction of catalytic performance. To address this problem, we comprehensively screened a series of transition-metal atom doped Ag-based SAAs. This screening process involves regression machine learning (ML) algorithms and a compressed-sensing data-analytics approach parameterized with density-functional inputs. The results demonstrate that Cu/Ag and Zn/Ag can efficiently activate and hydrogenate NO with small Φmax(η), a grand-canonical adaptation of the Gmax(η) descriptor, and exhibit higher affinity to NO over H adatoms to suppress the competing hydrogen evolution reaction. The NH3 selectivity is mainly determined by the s orbitals of the doped single-atom near the Fermi level. The catalytic activity of SAAs is highly correlated with the local environment of the active site. We further quantified the relationship between the intrinsic features of these active sites and Φmax(η). Our work clarifies the mechanism of NORR to NH3 and offers a design principle to guide the screen of highly active SAA catalysts.

Prediction of Ca2+ Binding Site in Proteins With a Fast and Accurate Method Based on Statistical Mechanics and Analysis of Crystal Structures.

Predicting the precise locations of metal binding sites within metalloproteins is a crucial challenge in biophysics. A fast, accurate, and interpretable computational prediction method can complement the experimental studies. In the current work, we have developed a method to predict the location of Ca2+ ions in calcium-binding proteins using a physics-based method with an all-atom description of the proteins, which is substantially faster than the molecular dynamics simulation-based methods with accuracy as good as data-driven approaches. Our methodology uses the three-dimensional reference interaction site model (3D-RISM), a statistical mechanical theory, to calculate Ca2+ ion density around protein structures, and the locations of the Ca2+ ions are obtained from the density. We have taken previously used datasets to assess the efficacy of our method as compared to previous works. Our accuracy is 88%, comparable with the FEATURE program, one of the well-known data-driven methods. Moreover, our method is physical, and the reasons for failures can be ascertained in most cases. We have thoroughly examined the failed cases using different structural and crystallographic measures, such as B-factor, R-factor, electron density map, and geometry at the binding site. It has been found that x-ray structures have issues in many of the failed cases, such as geometric irregularities and dubious assignment of ion positions. Our algorithm, along with the checks for structural accuracy, is a major step in predicting calcium ion positions in metalloproteins.

Comparing Large-Eddy Simulation and Gaussian Plume Model to Sensor Measurements of an Urban Smoke Plume

The fast prediction of the extent and impact of accidental air pollution releases is important to enable a quick and informed response, especially in cities. Despite this importance, only a small number of case studies are available studying the dispersion of air pollutants from fires in a short distance (O(1 km)) in urban areas. While monitoring pollution levels in Southampton, UK, using low-cost sensors, a fire broke out from an outbuilding containing roughly 3000 reels of highly flammable cine nitrate film and movie equipment, which resulted in high values of PM2.5 being measured by the sensors approximately 1500 m downstream of the fire site. This provided a unique opportunity to evaluate urban air pollution dispersion models using observed data for PM2.5 and the meteorological conditions. Two numerical approaches were used to simulate the plume from the transient fire: a high-fidelity computational fluid dynamics model with large-eddy simulation (LES) embedded in the open-source package OpenFOAM, and a lower-fidelity Gaussian plume model implemented in a commercial software package: the Atmospheric Dispersion Modeling System (ADMS). Both numerical models were able to quantitatively reproduce consistent spatial and temporal profiles of the PM2.5 concentration at approximately 1500 m downstream of the fire site. Considering the unavoidable large uncertainties, a comparison between the sensor measurements and the numerical predictions was carried out, leading to an approximate estimation of the emission rate, temperature, and the start and duration of the fire. The estimation of the fire start time was consistent with the local authority report. The LES data showed that the fire lasted for at least 80 min at an emission rate of 50 g/s of PM2.5. The emission was significantly greater than a ‘normal’ house fire reported in the literature, suggesting the crucial importance of the emission estimation and monitoring of PM2.5 concentration in such incidents. Finally, we discuss the advantages and limitations of the two numerical approaches, aiming to suggest the selection of fast-response numerical models at various compromised levels of accuracy, efficiency and cost.

Fast flow field prediction based on E(2)-equivariant steerable convolutional neural networks

In the field of flow field reconstruction, traditional deep learning models predominantly rely on standard convolutions, but their predictive accuracy remains limited. To address this issue, we explore the potential of E(2)-equivariant convolutions to enhance the predictive accuracy of deep learning models for fast flow field prediction. Unlike conventional convolutions, E(2)-equivariant convolutions offer a richer representation capability by better capturing geometric and structural information. Our neural network integrates an attention mechanism that leverages the signed distance function (SDF) to encode geometric details and an indicator matrix to incorporate boundary conditions. The model predicts velocity and pressure fields as outputs. We conducted experiments specifically targeting non-uniform steady laminar flows, and the results show a 16.1% reduction in overall error compared to models based on traditional convolutions while maintaining high efficiency. These findings indicate that E(2)-equivariant convolution, coupled with an attention mechanism, significantly improves flow field prediction by focusing on critical information and better representing complex geometries.

An improved permeability estimation model using integrated approach of hybrid machine learning technique and shapley additive explanation

Accurate reservoir permeability determination is crucial in hydrocarbon exploration and production. Conventional methods relying on empirical correlations and assumptions often result in high costs, time consumption, inaccuracies, and uncertainties. This study introduces a novel hybrid machine learning approach to predict the permeability of the Wangkwar formation in the Gunya oilfield, Northwestern Uganda. The group method of data handling with differential evolution (GMDH-DE) algorithm was used to predict permeability due to its capability to manage complex, nonlinear relationships between variables, reduced computation time, and parameter optimization through evolutionary algorithms. Using 1953 samples from Gunya-1 and Gunya-2 wells for training and 1563 samples from Gunya-3 for testing, the GMDH-DE outperformed the group method of data handling (GMDH) and random forest (RF) in predicting permeability with higher accuracy and lower computation time. The GMDH-DE achieved an R2 of 0.9985, RMSE of 3.157, MAE of 2.366, and ME of 0.001 during training, and for testing, the ME, MAE, RMSE, and R2 were 1.3508, 12.503, 21.3898, and 0.9534, respectively. Additionally, the GMDH-DE demonstrated a 41% reduction in processing time compared to GMDH and RF. The model was also used to predict the permeability of the Mita Gamma well in the Mandawa basin, Tanzania, which lacks core data. Shapley additive explanations (SHAP) analysis identified thermal neutron porosity (TNPH), effective porosity (PHIE), and spectral gamma-ray (SGR) as the most critical parameters in permeability prediction. Therefore, the GMDH-DE model offers a novel, efficient, and accurate approach for fast permeability prediction, enhancing hydrocarbon exploration and production.

LKFlowNet: A deep neural network based on large kernel convolution for fast and accurate nonlinear fluid-changing prediction

LKFlowNet: A deep neural network based on large kernel convolution for fast and accurate nonlinear fluid-changing prediction

Application of machine learning technique for a fast forecast of aggregation kinetics in space-inhomogeneous systems

Modelling of aggregation processes in space-inhomogeneous systems is extremely numerically challenging since complicated aggregation equations, Smoluchowski equations, are to be solved at each space point along with the computation of particle propagation. Low rank approximation for the aggregation kernels can significantly speed up the solution of Smoluchowski equations, while the particle propagation could be done in parallel. Yet the numerics with many aggregate sizes remains quite resource-demanding. Here, we explore the way to reduce the amount of direct computations by replacing the actual numerical solution of the Smoluchowski equations with the respective density transformations learned with the application of one of machine learning (ML) methods, the conditional normalising flow. We demonstrate that the ML predictions for the space distribution of aggregates and their size distribution require drastically shorter computation time and agree fairly well with the results of direct numerical simulations. Such an opportunity of a quick forecast of space-dependent particle size distribution could be important in practice, especially for the fast (on the timescale of data reading) prediction and visualisation of pollution processes, providing a tool with a reasonable trade off between the prediction accuracy and the computational time.

Prediction of teicoplanin plasma concentration in critically ill patients: a combination of machine learning and population pharmacokinetics.

Teicoplanin has been widely used in patients with infections caused by Staphylococcus aureus, especially for critically ill patients. The pharmacokinetics (PK) of teicoplanin vary between individuals and within the same individual. We aim to establish a prediction model via a combination of machine learning and population PK (PPK) to support personalized medication decisions for critically ill patients. A retrospective study was performed incorporating 33 variables, including PPK parameters (clearance and volume of distribution). Multiple algorithms and Shapley additive explanations were employed for feature selection of variables to determine the strongest driving factors. The performance of each algorithm with PPK parameters was superior to that without PPK parameters. The composition of support vector regression, categorical boosting and a backpropagation neural network (7:2:1) with the highest R2 (0.809) was determined as the final ensemble model. The model included 15 variables after feature selection, of which the predictive performance was superior to that of models considering all variables or using only PPK. The R2, mean absolute error, mean squared error, absolute accuracy (±5 mg/L) and relative accuracy (±30%) of external validation were 0.649, 3.913, 28.347, 76.12% and 76.12%, respectively. Our study offers a non-invasive, fast and cost-effective prediction model of teicoplanin plasma concentration in critically ill patients. The model serves as a fundamental tool for clinicians to determine the effective plasma concentration range of teicoplanin and formulate individualized dosing regimens accordingly.