Dense Neural Network Research Articles

RETRACTED ARTICLE: Down syndrome markers classification via dense neural network in ultrasound foetal image

Down syndrome is a genetically born disorder among infants that occurs during the development of the foetus. Trisomy 21, a chromosome imbalance disorder is a leading cause of the down syndrome. Numerous machine learning (ML) models have been used to identify down syndrome in ultrasound images of foetuses, but the development of deep learning, offers an enormous advantage over ML models in accuracy. However, the existing models have focused on down syndrome as a nasal bone (NB) length or nuchal translucency (NT). In this paper, a novel automatic dense convolution neural network (DConN) is proposed to isolate and measure the down syndrome marker particularly NB length and NT. It is necessary to extract texture features precisely from ultrasound images to classify them accurately. Initially, the test image is processed using an anisotropic diffusion filter to remove the noise. The pre-processed foetal images are subjected to region of interest extraction to isolate the Down syndrome makers from the ultrasound image. The proposed DConN is used to classify the normal and abnormal foetal based on NT and NB length of the foetal in US images. The performance metrics namely sensitivity, accuracy, specificity, F1 score, and precision are considered for validating the effectiveness of the proposed model. From the experimental the proposed DconN achieves the overall accuracy of 99.01%. The proposed method improves the overall accuracy of 3.9%, 1.6% and 0.41% better than cascaded ML, SIFT + GRNN and Modified AdaBoost, respectively.

C-3 | A Machine Learning Algorithm to Detect Right Heart Failure in Cardiogenic Shock when Using Impella

Right heart failure (RHF) is common amongst patients in cardiogenic shock (CS) and best diagnosed with use of a pulmonary artery catheter (PAC). Unfortunately use of PACs remains low (<20%) even in cases of CS. Early identification and treatment of RVF may improve outcomes for such patients. We propose a machine learning algorithm to detect RVF in patients presenting with cardiogenic shock who are treated with an Impella CP. We leveraged clinical data collected in the National Cardiogenic Shock Initiative (NCSI) to develop and train a RHF Indicator model: RHFNet, which is a dense neural network model that can infer the likelihood of RHF from Impella CP data (Figure). The model was developed using 129 NCSI patients (N=129 development subjects: 49 (32%) RHF, 80 (68%) No RHF) on the data from first 24 hours of Impella CP support. For validation, 84 independent cases were identified from the Abiomed IQ database that had concurrent CVP and PAPI measurements. These cases were assigned to a cohort based on their CVP and PAPI values to create a validation set. RHFNet was trained on a 66% random sample of development subjects (N=86 training subjects) and tested on the remaining 34% subjects (N=43 test subjects) to establish the viability of the solution. The model achieved a mean AUC score of 0.80 on the test subjects. On independent IQ cases (N=84 validation subjects), RHFNet achieved a similar AUC score of 0.78. RHF can be identified using only signals from a left-sided MCS device (Impella CP).

Function of CD8+, conventional CD4+, and regulatory CD4+ T cell identification in lung cancer

Lung cancer is the malignant tumor with the highest mortality rate in the world. There is obvious heterogeneity within the tumor. Single cell sequencing technology enables scholars to obtain information about the cell type, status, subpopulation distribution and communication behavior between cells in the tumor microenvironment from the cellular level. However, due to the problem of sequencing depth, some genes with low expression cannot be detected, which results in that most of the specific genes of immune cells cannot be recognized, and lead to defects in the functional identification of immune cells. In this paper, we used single cell sequencing data of 12346 T cells in 14 treatment-naïve non-small-cell lung cancer patients to identify immune cell-specific genes and infer the function of three types of T cells. The method, named GRAPH-LC, implemented this function by gene interaction network and graph learning methods. Graph learning methods are used to extract genes feature and dense neural network is used to identify immune cell-specific genes. The experiments on 10-cross validation shows that the AUROC and AUPR reached at least 0.802, 0.815 on identifying cell-specific genes of three types of T cells. And we did functional enrichment analysis on the top 15 expressed genes. By functional enrichment analysis, we got 95 GO terms and 39 KEGG pathways that related to three types of T cells. The use of this technology will help to deeply understand the mechanism of the occurrence and development of lung cancer, find new diagnostic markers and therapeutic targets, and provide a theoretical reference for the precise treatment of lung cancer patients in the future.

Automatic segmentation and fibre orientation estimation from low resolution X-ray computed tomography images of 3D woven composites

In this work, we compare different methods for the automatic segmentation of microtomographic images at the mesoscopic scale of 3D woven carbon fibre composite materials. We use a dense neural network approach fed with features based on the monogenic signal and features learned by a convolutional neural network. To determine local fibre orientations in the composite, we also propose to use the monogenic signal with a structure tensor approach. The performance of the segmentation methods is compared on a first sample for which a reference solution is given. To demonstrate the global capability of the method, the neural networks are then used to segment a second similar sample, for which there is no reference solution. A transfer learning strategy is proposed to take advantage of the knowledge learned on the first sample and to limit the interaction with the user. This simply consists of re-learning only the decision layers, thus limiting the number of examples required. Finally, the best result is obtained with a combination of the signal-based approach and the learning-based approach by merging the monogenic and convolutional neural network features.

Detecting Dravidian Offensive Posts in MIoT: A Hybrid Deep Learning Framework

Hate speech and Offensive Posts (OP) detection on Smart Multimedia Internet of Things (MIoT) have been an active issue for researchers. MIoT media texts in non-native English-speaking countries are often code-mixed or script mixed/switched. This paper proposes an ensemble-based Deep Learning (DL) framework comprised of a Convolutional Neural Network (CNN) and a Dense Neural Network (DNN) for identifying hate and OP in Malayalam Code-Mixed (MCM), Tamil Code-Mixed (TCM), and Malayalam Script-Mixed (MSM) MIoT media postings. Word-level and character-level features are utilized in the convolutional neural network. In contrast, the dense neural network uses character-level Term Frequency-Inverse Document Frequency (TF-IDF) features. The inclusion of character-level features in the proposed ensemble framework resulted in state-of-the-art performance for TCM and MCM datasets, with weighted F 1 -score of 0.91 and 0.78, respectively, and comparable performance for MSM posts, with a weighted F 1 -score of 0.95.

Machine Learning to Predict Three Types of Outcomes After Traumatic Brain Injury Using Data at Admission: A Multicenter Study for Development and Validation.

The difficulty of accurately identifying patients who would benefit from promising treatments makes it challenging to prove the efficacy of novel treatments for traumatic brain injury (TBI). Although machine learning is being increasingly applied to this task, existing binary outcome prediction models are insufficient for the effective stratification of TBI patients. The aim of this study was to develop an accurate 3-class outcome prediction model to enable appropriate patient stratification. To this end, retrospective balanced data of 1200 blunt TBI patients admitted to six Japanese hospitals from January 2018 onwards (200 consecutive cases at each institution) were used for model training and validation. We incorporated 21 predictors obtained in the emergency department, including age, sex, six clinical findings, four laboratory parameters, eight CT findings, and an emergency craniotomy. We developed two machine learning models (XGBoost and dense neural network) and logistic regression models to predict 3-class outcomes based on the Glasgow Outcome Scale-Extended (GOSE) at discharge. The prediction models were developed using a training dataset with n = 1000, and their prediction performances were evaluated over two validation rounds on a validation dataset (n = 80) and a test dataset (n = 120) using the bootstrap method. Of the 1200 patients in aggregate, the median patient age was 71 years, 199 (16.7%) exhibited severe TBI, and emergency craniotomy was performed on 104 patients (8.7%). The median length of stay was 13.0 days. The 3-class outcomes were good recovery/moderate disability for 709 patients (59.1%), severe disability/vegetative state in 416 patients (34.7%), and death in 75 patients (6.2%). XGBoost model performed well with 69.5% sensitivity, 82.5% accuracy, and an area under the receiver operating characteristic curve of 0.901 in the final validation. In terms of the receiver operating characteristic curve analysis, the XGBoost outperformed the neural network-based and logistic regression models slightly. In particular, XGBoost outperformed the logistic regression model significantly in predicting severe disability/vegetative state. Although each model predicted favorable outcomes accurately, they tended to miss the mortality prediction. The proposed machine learning model was demonstrated to be capable of accurate prediction of in-hospital outcomes following TBI, even with the three GOSE-based categories. As a result, it is expected to be more impactful in the development of appropriate patient stratification methods in future TBI studies than conventional binary prognostic models. Furthermore, outcomes were predicted based on only clinical data obtained from the emergency department. However, developing a robust model with consistent performance in diverse scenarios remains challenging, and further efforts are needed to improve generalization performance.

Deep learning health management diagnostics applied to the NIST smoke experiments

Fire is one of the most important hazards that must be considered in advanced nuclear power plant safety assessments. The Nuclear Regulatory Commission (NRC) has developed a large collection of experimental data and associated analyses related to the study of fire safety. In fact, computational fire models are based on quantitative comparisons to those experimental data. During the modeling process, it is important to develop diagnostic health management systems to check the equipment status in fire processes. For example, a fire sensor does not directly provide accurate and complex information that nuclear power plants (NPPs) require. With the assistance of the machine learning method, NPP operators can directly get information on local, ignition, fire material of an NPP fire, instead of temperature, smoke obscuration, gas concentration, and alarm signals. In order to improve the predictive capabilities, this work demonstrates how the deep learning classification method can be used as a diagnostic tool in a specific set of fire experiments. Through a single input from a sensor, the deep learning tool can predict the location and type of fire. This tool also has the capability to provide automatic signals to potential passive fire safety systems. In this work, test data are taken from a specific set of the National Institute of Standards and Technology (NIST) fire experiments in a residential home and analyzed by using the machine learning classification models. The networks chosen for comparison and evaluation are the dense neural networks, convolutional neural networks, long short-term memory networks, and decision trees. The dense neural network and long short-term memory network produce similar levels of accuracy, but the convolutional neural network produces the highest accuracy.

Deep neural networks and gradient-weighted class activation mapping to classify and analyze EEG

Deep Neural Networks (DNN) have significantly improved the capabilities for analysis and classification of data, including that of biomedical signals (eg. ElectroencephalogramEEG). Optimal classification of EEG signals from seizure onset zones has been challenging especially given the complexity of signals arising from multiple locations. Also, underlying electrophysiological abnormalities which signify epileptogenic zones have not been clearly defined. Previous studies have demonstrated, automatic feature generation based on deep learning as a useful tool for interictal epileptiform discharge (IEDs) detection. Also signals with transformations have been used in convolutional neural network (CNN) based models in the past for classifying EEG data. We explored the use of deep learning for identification of the seizure onset zones using regular dense neural network and CNN based models. After the training the model using sample data the results were validated with a smaller percentage (10%) of the data. The models were noted to be accurate in predicting the seizure onset zones with significant degree of accuracy (87–99%) with a much shorter duration of signal recorded, compared to previous studies. We further investigated model’s decision-making process with heatmapping (gradient-weighted class activation map: Grad-CAM) approach, combining with signal processing using Hilbert transform. To identify the visible features in the signal, which maximally contribute to the seizure onset zone prediction, correlation of heatmap and analytical signal of the EEG, was calculated. A high correlation between heatmap and analytical signal was noted, suggesting that the model may be utilizing the higher power regions of the data in decision making process. This study demonstrates the potential use of DNN based strategies in identifying the epileptogenic zones in the intracranial EEG. It also shows the heatmapping strategies can help in establishing how specific signal patterns may be contributing to the decision making of DNN.

Machine learning-based prediction of operation conditions from plasma plume images of atmospheric-pressure plasma reactors

A technique was proposed in this paper to monitor the key operating conditions of a plasma abatement system, which are the concentration of the carbon-containing process gas and the treatment flowrate, from a plasma plume image acquired using an inexpensive color camera. The technique is based on the observation that the shape and color of the plasma plume vary with the variations in the specific energy input and plasma gas composition. In addition, because these variations are marginal and it is challenging to identify an analytical relationship between these variations and the operating conditions, the prediction model is obtained in a data-driven manner. Specifically, the model was composed of a set of convolutional autoencoders (CAEs) and a dense neural network. Furthermore, it was trained only with images captured under normal operation so that (1) images captured under abnormal operations could be identified based on the reconstruction error of the trained CAEs and (2) predictions are made only on normal images. As a demonstration, methane was tested as a process gas, and oxygen was used as a reaction agent in a nitrogen-rich environment. The test results showed that the optimized model could predict the treatment flowrate and process gas concentration with 96% probability within ±3.08 slpm and ±300 ppm, respectively.

Predicting lake water quality index with sensitivity-uncertainty analysis using deep learning algorithms

Regular monitoring and assessment of water quality is essential to maintain the quality of lake water. A commonly used method for assessing water quality is the Water Quality Index (WQI). Artificial intelligence (AI) can predict the WQI rapidly and more accurately than conventional methods. In this study, a stacking ensemble model based on Deep Learning (DL) was proposed to predict the WQI by integrating three models: Gradient Boosting Machine (GBM), Generalized Linear Model (GLM) and Neural Network (NN). The performance of this model was compared with that of Deep Dense Neural Network (DNN) and Convolutional Neural Network (CNN). The inclusion of a DNN model, which was used for the first time in water pollution research to perform sensitivity and uncertainty analysis in predicting WQI, added a new dimension to the workflow. The mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE) and correlation coefficient (R2) were used to evaluate the performance of the WQI prediction models. The results showed that the WQI of the water samples ranged from 90.75 to 145.29, indicating poor to very poor water quality. The validation results showed that the stack model was superior to all other models in terms of prediction accuracy, with MSE of 25.77, RMSE of 5.07, MAE of 3.5 and R2 of 0.98. The DNN-based sensitivity analysis showed that pH and turbidity significantly affect the WQI and should be monitored to minimise water pollution. The uncertainty analysis indicated that electrical conductivity and total dissolved solids have the highest uncertainty in predicting WQI. This study provides researchers, decision makers and water scientists with accurate information on water quality and enables the implementation of clean production processes to reduce pollution and improve water quality.

Modeling early-universe energy injection with dense neural networks

We show that dense neural networks can be used to accurately model the cooling of high-energy particles in the early universe, in the context of the public code package darkhistory. darkhistory self-consistently computes the temperature and ionization history of the early universe in the presence of exotic energy injections, such as might arise from the annihilation or decay of dark matter. The original version of darkhistory uses large precomputed transfer function tables to evolve photon and electron spectra in redshift steps, which require a significant amount of memory and storage space. We present a light version of darkhistory that makes use of simple dense neural networks to store and interpolate the transfer functions, reducing their total size by a factor of $\ensuremath{\sim}400$, from $\ensuremath{\sim}17\text{ }\text{ }\mathrm{Gb}$ to $\ensuremath{\sim}45\text{ }\text{ }\mathrm{Mb}$, which performs well on small computers with limited memory and storage. This method anticipates future expansion with additional parametric dependence in the transfer functions without requiring exponentially larger data tables.

Inferring influence of people's emotions at court on defendant's emotions using a prediction model.

People's emotions may be affected by the sound environment in court. A courtroom's sound environment usually consists of the people's voices, such as the judge's voice, the plaintiff's voice, and the defendant's voice. The judge, plaintiff, and defendant usually express their emotions through their voices. Human communication is heavily reliant on emotions. Emotions may also reflect a person's condition. Therefore, People's emotions at the Court must be recognized, especially for vulnerable groups, and the impact of the sound on the defendant's motions and judgment must be inferred. However, people's emotions are difficult to recognize in a courtroom. In addition, as far as we know, no existing study deals with the impact of sound on people in court. Based on sound perception, we develop a deep neural network-based model to infer people's emotions in our previous work. In the proposed model, we use the convolutional neural network and long short-term memory network to obtain features from speech signals and apply a dense neural network to infer people's emotions. Applying the model for emotion prediction based on sound at court, we explore the impact of sound at court on the defendant. Using the voice data collected from fifty trail records, we demonstrate that the voice of the judge can affect the defendant's emotions. Angry, neutrality and fear are the top three emotions of the defendant in court. In particular, the judge's voice expressing anger usually induces fear in the defendant. The plaintiff's angry voice may not have a substantial impact on the defendant's emotions.

Classification of Ayurveda constitution types: a deep learning approach

In precision medicine, there has been a paradigm shift, from a population average based one-size-fits all, to an individualized approach. The aim is to tailor a medicine according to an individual’s genetic make-up and response to drug, diet, and environment. Machine learning algorithms are being developed to dissect inter-individual variability in the population. This is more important for prevention of diseases in healthy individuals and early actionable interventions. In Ayurveda, the basic tenet of the practice is personalized—an individual’s constitution type “Prakriti” is used for predicting healthy and disease trajectories as well as individualized management. Prakriti refers to the nature of the body in terms of doshas, and the time of conception is decided according to the predominance of doshas. It does not change throughout life and is responsible for a person’s physical and mental characteristics. Any population can be stratified into seven Prakriti groups amongst which the Vata, Pitta and Kapha are the three phenotypic extremes with contrasting disease susceptibilities and response. Prakriti is assessed through examination of multi-system attributes and has been formalized into a questionnaire that captures 120 feature attributes for objective assessment. We use for the first time—powerful dense neural network deep learning algorithms for predicting Prakriti classes. The study was carried out on a validated dataset of 233 extreme Prakriti individuals derived from two genetically diverse cohorts from North and West India. We benchmark our method with the state-of-the-art supervised machine learning and establish a learning approach for transferring features from the Western to Northern Indian data and vice versa as well. Our study provides a computational framework for predicting Prakriti which can be implemented in large scale studies of precision medicine. The dataset and the baseline implementation can be found in ( https://github.com/skarifahmed/DeepAyurveda )

Pulse shape simulation and discrimination using machine learning techniques

An essential metric for the quality of a particle-identification experiment is its statistical power to discriminate between signal and background. Pulse shape discrimination (PSD) is a basic method for this purpose in many nuclear, high-energy and rare-event search experiments where scintillation detectors are used. Conventional techniques exploit the difference between decay-times of the pulses from signal and background events or pulse signals caused by different types of radiation quanta to achieve good discrimination. However, such techniques are efficient only when the total light-emission is sufficient to get a proper pulse profile. This is only possible when adequate amount of energy is deposited from recoil of the electrons or the nuclei of the scintillator materials caused by the incident particle on the detector. But, rare-event search experiments like direct search for dark matter do not always satisfy these conditions. Hence, it becomes imperative to have a method that can deliver a very efficient discrimination in these scenarios. Neural network based machine-learning algorithms have been used for classification problems in many areas of physics especially in high-energy experiments and have given better results compared to conventional techniques. We present the results of our investigations of two network based methods viz. Dense Neural Network and Recurrent Neural Network, for pulse shape discrimination and compare the same with conventional methods.

PSO-Based Ensemble Meta-Learning Approach for Cloud Virtual Machine Resource Usage Prediction

To meet the increasing demand for its services, a cloud system should make optimum use of its available resources. Additionally, the high and low oscillations in cloud workload are another significant symmetrical issue that necessitates consideration. A suggested particle swarm optimization (PSO)-based ensemble meta-learning workload forecasting approach uses base models and the PSO-optimized weights of their network inputs. The proposed model employs a blended ensemble learning strategy to merge three recurrent neural networks (RNNs), followed by a dense neural network layer. The CPU utilization of GWA-T-12 and PlanetLab traces is used to assess the method’s efficacy. In terms of RMSE, the approach is compared to the LSTM, GRU, and BiLSTM sub-models.

GPCRLigNet: rapid screening for GPCR active ligands using machine learning.

Molecules with bioactivity towards G protein-coupled receptors represent a subset of the vast space of small drug-like molecules. Here, we compare machine learning models, including dilated graph convolutional networks, that conduct binary classification to quickly identify molecules with activity towards G protein-coupled receptors. The models are trained and validated using a large set of over 600,000 active, inactive, and decoy compounds. The best performing machine learning model, dubbed GPCRLigNet, was a surprisingly simple feedforward dense neural network mapping from Morgan fingerprints to activity. Incorporation of GPCRLigNet into a high-throughput virtual screening workflow is demonstrated with molecular docking towards a particular G protein-coupled receptor, the pituitary adenylate cyclase-activating polypeptide receptor type 1. Through rigorous comparison of docking scores for molecules selected with and without using GPCRLigNet, we demonstrate an enrichment of potentially potent molecules using GPCRLigNet. This work provides a proof of principle that GPCRLigNet can effectively hone the chemical search space towards ligands with G protein-coupled receptor activity.

A comparative study of machine learning and deep learning algorithms for padel tennis shot classification

Data processing in sports is a phenomenon increasingly present at all levels, from professionals in search of tools to improve their performance to beginners motivated by the quantification of their physical activity. In this work, a comparison between some of the main machine learning and deep learning algorithms is carried out in order to classify padel tennis strokes. Up to 13 representative padel tennis strokes are classified. Before a classification of padel tennis strokes is performed, a sufficiently representative data set is needed that collects numerous examples of the performance of these strokes. Since there was no similar data set in the literature, we proceeded to the creation of such a data set, for which we developed a data collection system based on an electronic device with an inertial measurement unit. Using the developed data set, the machine learning and deep learning algorithms were hyperparameterized to compare their performance under the best possible configurations. The algorithms were fed with both the temporal series of the acceleration and speed of the six degrees of freedom and also with feature engineering input, consisting in calculating the mean, maximum, and minimum values for each axis. The algorithms evaluated are: fully connected or dense neural networks, 1D convolutional neural networks, decision tree, K nearest neighbors, support vector machines, and eigenvalue classification. According to the results achieved, the best algorithm is the 1D convolutional neural network with temporal series input that achieves an accuracy higher than 93%. However, other simpler algorithms such as dense networks and support vector machines achieve similar results.

Enhancing Logistic Regression Using Neural Networks for Classification in Actuarial Learning

We developed a methodology for the neural network boosting of logistic regression aimed at learning an additional model structure from the data. In particular, we constructed two classes of neural network-based models: shallow–dense neural networks with one hidden layer and deep neural networks with multiple hidden layers. Furthermore, several advanced approaches were explored, including the combined actuarial neural network approach, embeddings and transfer learning. The model training was achieved by minimizing either the deviance or the cross-entropy loss functions, leading to fourteen neural network-based models in total. For illustrative purposes, logistic regression and the alternative neural network-based models we propose are employed for a binary classification exercise concerning the occurrence of at least one claim in a French motor third-party insurance portfolio. Finally, the model interpretability issue was addressed via the local interpretable model-agnostic explanations approach.

Tailored magnetic resonance fingerprinting

Neuroimaging of certain pathologies requires both multi-parametric qualitative and quantitative imaging. The role of the quantitative MRI (qMRI) is well accepted but suffers from long acquisition times leading to patient discomfort, especially in geriatric and pediatric patients. Previous studies show that synthetic MRI can be used in order to reduce the scan time and provide qMRI as well as multi-contrast data. However, this approach suffers from artifacts such as partial volume and flow. In order to increase the scan efficiency (the number of contrasts and quantitative maps acquired per unit time), we designed, simulated, and demonstrated rapid, simultaneous, multi-contrast qualitative (T1 weighted, T1 fluid attenuated inversion recovery (FLAIR), T2 weighted, water, and fat), and quantitative imaging (T1 and T2 maps) through the approach of tailored MR fingerprinting (TMRF) to cover whole-brain in approximately four minutes.We performed TMRF on in vivo four healthy human brains and in vitro ISMRM/NIST phantom and compared with vendor supplied gold standard (GS) and MRF sequences. All scans were performed on a 3 T GE Premier system and images were reconstructed offline using MATLAB. The reconstructed qualitative images were then subjected to custom DL denoising and gradient anisotropic diffusion denoising. The quantitative tissue parametric maps were reconstructed using a dense neural network to gain computational speed compared to dictionary matching. The grey matter and white matter tissues in qualitative and quantitative data for the in vivo datasets were segmented semi-automatically. The SNR and mean contrasts were plotted and compared across all three methods. The GS images show better SNR in all four subjects compared to MRF and TMRF (GS > TMRF>MRF). The T1 and T2 values of MRF are relatively overestimated as compared to GS and TMRF. The scan efficiency for TMRF is 1.72 min−1 which is higher compared to GS (0.32 min−1) and MRF (0.90 min−1).

Application of dense neural networks for manifold-based modeling of flame-wall interactions

Artifical neural networks (ANNs) are universal approximators capable of learning any correlation between arbitrary input data with corresponding outputs, which can also be exploited to represent a low-dimensional chemistry manifold in the field of combustion. In this work, a procedure is developed to simulate a premixed methane-air flame undergoing side-wall quenching utilizing an ANN chemistry manifold. In the investigated case, the flame characteristics are governed by two canonical problems: the adiabatic flame propagation in the core flow and the non-adiabatic flame- wall interaction governed by enthalpy losses to the wall. Similar to the tabulation of a Quenching Flamelet-Generated Manifold (QFM), the neural network is trained on a 1D head-on quenching flame database to learn the intrinsic chemistry manifold. The control parameters (i.e. the inputs) of the ANN are identified from thermo-chemical state variables by a sparse principal component analysis (PCA) without using prior knowledge about the flame physics. These input quantities are then transported in the coupled CFD solver and used for manifold access during simulation runtime. The chemical source terms are corrected at the manifold boundaries to ensure bounded- ness of the thermo-chemical state at all times. Finally, the ANN model is assessed by comparison to simulation results of the 2D side-wall quenching (SWQ) configuration with detailed chemistry and with a flamelet-based manifold (QFM).