Three-layer Feedforward Neural Network Research Articles

Precision Regulation and Forecasting of Greenhouse Tomato Growth Conditions Using an Improved GA-BP Model

Tomatoes are annual herbaceous plants of the family Solanaceae. They have very stringent requirements for their growing environment and climate conditions. To precisely control the greenhouse environment for tomato growth, this project designed and implemented a monitoring system utilizing programmable logic controllers and a data acquisition system for monitoring. Sensors are installed at key locations in the greenhouse, such as near the ground, in the plant canopy, and under the roof, to monitor environmental parameters such as temperature, humidity, light intensity, and carbon dioxide concentration in real time. A three-layer feedforward GA-BP neural network model was established using soil temperature, soil humidity, air humidity, and illumination as predictive parameters. This technology predicts optimal environmental parameters and performs real-time monitoring and intelligent control, overcoming the traditional BP neural network’s drawbacks of slow convergence, susceptibility to disturbance, and poor generalization ability, with the current model’s average prediction error being less than 5%. It provides a theoretical basis and decision support for precision control and related environmental forecasting in agricultural projects. Based on the growth characteristics, physiological and morphological features of tomatoes, and the comprehensive impact of planting conditions, the technology enhances tomato yield and quality while reducing production costs and labor inputs. It mainly includes the establishment of tomato growth models, measurement and control of environmental parameters, and the design and optimization of intelligent decision control systems. Research outcomes in this field are significant for improving agricultural production efficiency, ensuring food security, and promoting sustainable agricultural development.

Enhanced predictive optimization of methane dry reforming via ResponseSurface methodology and artificial neural network approaches: Insights using a novel nickel-strontium-zirconium-aluminum catalyst

This study investigates the molecular dynamics of methane dry reforming catalyzed by a novel nickel-strontium-zirconium-aluminum (5Ni+3Sr/10Zr+Al) catalyst, leveraging both Response Surface Methodology (RSM) and Radial Basis Function Neural Network (RBFNN) for predictive optimization. Focusing on the impact of operational parameters—hourly space velocity, reaction temperature, and CO2:CH4 mole ratio—on the conversion rates and formation of reaction components, we aim to predict optimal conditions and corresponding process variables. Through a comparison of a three-layer Feed Forward Neural Network, optimized at a 3:10:1 topology, with traditional RSM approaches, our findings highlight the superior predictive capabilities of ANN models. Notably, ANN demonstrated exceptional performance with Radj2and F_Ratio values significantly surpass those of RSM, alongside lower statistical error terms. This superiority is attributed to ANN's robust handling of nonlinear relationships between inputs and outputs, asserting its potential for enhancing predictive accuracy in chemical process optimization. At optimum predicted conditions like 1 CH4/CO2,750 °C reaction temperature, 12000 cm3g−1h−1 space velocity, NiSrZrAl outperformed with > 85 % CH4 and CO2 conversion with H2/CO ∼1 up to 20 h time on stream. Our research underscores the importance of integrating advanced modeling techniques for the efficient and accurate prediction of catalytic reactions, offering valuable insights for future applications in chemical engineering and catalysis.

Performance Analysis of Waste Biomass Gasification and Renewable Hydrogen Production by Neural Network Algorithm

Abstract This study assesses renewable hydrogen production via gasification of residual biomass, using artificial neural networks (ANNs) for predictive modeling. The process uses residues from sugarcane and orange harvests, sewage sludge, corn byproducts, coffee remnants, eucalyptus remains, and urban waste. Simulation data from aspen plus® software predict hydrogen conversion from each biomass type, with a three-layer feedforward neural network algorithm used for model construction. The model showed high accuracy, with R2 values exceeding 0.9941 and 0.9931 in training and testing datasets, respectively. Performance metrics revealed a maximum higher heating value (HHV) of 18.1 MJ/kg for sewage sludge, the highest cold gas efficiency for urban and orange waste (82.2% and 80.6%), and the highest carbon conversion efficiency for sugarcane bagasse and orange residue (92.8% and 91.2%). Corn waste and sewage sludge yielded the highest hydrogen mole fractions (0.55 and 0.52). The system can reach relative exergy efficiencies from 24.4% for sugarcane straw residues to 42.6% for sugarcane bagasse. Rational exergy efficiencies reached from 23.7% (coffee waste) to 39.0% (sugarcane bagasse). This research highlights the potential of ANNs in forecasting hydrogen conversion and assessing the performance of gasification-based renewable hydrogen procedures using biomass wastes.

Reaction Rates of OH Radicals with CH3OCF2CHFCF3 and CHF2CF2OCH2CF2CHF2: Measurements and Estimation Using Neural Network Method.

The reaction rates of OH radicals with CH3OCF2CHFCF3 (k1) and CHF2CF2OCH2CF2CHF2 (k2) were measured over a temperature range of 250-430 K. Kinetic measurements were performed using the flash and laser photolysis methods combined with a laser-induced fluorescence technique. The Arrhenius rate parameters were determined as k1 = (2.52 ± 0.25) × 10-12·exp[-(1390 ± 30)/T], k2 = (1.83 ± 0.20) × 10-12·exp[-(1420 ± 35)/T] cm3 molecule-1 s-1. The infrared absorption spectra of the two hydrofluoroethers were measured at approximately 298 K in 760 Torr of N2. The atmospheric lifetimes of CH3OCF2CHFCF3 and CHF2CF2OCH2CF2CHF2 have been estimated as 2.5 and 3.8 years, respectively, and their global warming potentials were determined as 245 and 405, respectively. Additionally, a method, using a three-layered feed-forward neural network, for estimating the rates of the reaction of the OH radicals with alkanes, ethers, and alcohols was investigated. The ratios of the calculated reaction rates to the observed ones agreed within a factor of 2. The ability of the neural network method to predict reaction rates was examined by using a leave-one-out test.

Smart City Distributed Trade Economy Dynamic Scheduling of E-Commerce Platform Based on Reinforcement Learning Algorithm

Abstract In order to improve the reliability and economy of decentralized trade economy dynamic scheduling on e-Commerce platforms and shorten the running time of decentralized trade economy dynamic scheduling on e-Commerce platforms, a decentralized trade economy dynamic scheduling method based on the reinforcement learning algorithm is proposed. In this paper, we analyze the basic theory of the reinforcement learning algorithm, study the Q-learning algorithm, build a neural network to fit the value model, and initialize the reinforcement learning algorithm. With Markov decision process as the framework model, the optimal state behavior value function is updated by using the modeless discounted reward reinforcement learning algorithm Q-learning as the value iteration method. Gibbs distribution is used to construct exploratory random strategies to select behaviors with probability. Using the reinforcement learning algorithm and the three-layer feedforward neural network as the approximator of the state behavior value function, this paper studies the generalization of the value function faced by the decentralized trade economy dynamic scheduling of e-Commerce platforms and realizes the decentralized trade economy dynamic scheduling of e-Commerce platforms. The experimental results show that the proposed method can effectively improve the reliability and economy of the decentralized trade economy dynamic scheduling of e-Commerce platforms.

Hybrid deep modeling of a CHO-K1 fed-batch process: combining first-principles with deep neural networks.

Introduction: Hybrid modeling combining First-Principles with machine learning is becoming a pivotal methodology for Biopharma 4.0 enactment. Chinese Hamster Ovary (CHO) cells, being the workhorse for industrial glycoproteins production, have been the object of several hybrid modeling studies. Most previous studies pursued a shallow hybrid modeling approach based on three-layered Feedforward Neural Networks (FFNNs) combined with macroscopic material balance equations. Only recently, the hybrid modeling field is incorporating deep learning into its framework with significant gains in descriptive and predictive power. Methods: This study compares, for the first time, deep and shallow hybrid modeling in a CHO process development context. Data of 24 fed-batch cultivations of a CHO-K1 cell line expressing a target glycoprotein, comprising 30 measured state variables over time, were used to compare both methodologies. Hybrid models with varying FFNN depths (3-5 layers) were systematically compared using two training methodologies. The classical training is based on the Levenberg-Marquardt algorithm, indirect sensitivity equations and cross-validation. The deep learning is based on the Adaptive Moment Estimation Method (ADAM), stochastic regularization and semidirect sensitivity equations. Results and conclusion: The results point to a systematic generalization improvement of deep hybrid models over shallow hybrid models. Overall, the training and testing errors decreased by 14.0% and 23.6% respectively when applying the deep methodology. The Central Processing Unit (CPU) time for training the deep hybrid model increased by 31.6% mainly due to the higher FFNN complexity. The final deep hybrid model is shown to predict the dynamics of the 30 state variables within the error bounds in every test experiment. Notably, the deep hybrid model could predict the metabolic shifts in key metabolites (e.g., lactate, ammonium, glutamine and glutamate) in the test experiments. We expect deep hybrid modeling to accelerate the deployment of high-fidelity digital twins in the biopharma sector in the near future.

Continuous–Discrete Time Neural Network Observer for Nonlinear Dynamic Systems Application to Vehicle Systems

This paper proposes a novel continuous-discrete (sampled data) time neural network (NSNN) observer for nonlinear systems. It can therefore be applied to systems with a high degree of non-linearity with no prior knowledge of the system dynamics. The proposed observer is a three-layer feedforward neural network that has been intensively trained using the error backpropagation learning algorithm, which includes an e-modification term to ensure robustness of the observer. A structure of the output predictor with a corrective term is added in the structure of the NN observer to overcome the problem of discrete time measurement. Simulations using MATLAB and CarSim are illustrated to demonstrate the performance of the proposed state observer strategy to reconstruct the state variables and parameters of a vehicle system.

Real-Time Vehicle Classification System Using a Single Magnetometer

Vehicle count and classification data are very important inputs for intelligent transportation systems (ITS). Magnetic sensor-based technology provides a very promising solution for the measurement of different traffic parameters. In this work, a novel, real-time vehicle detection and classification system is presented using a single magnetometer. The detection, feature extraction, and classification are performed online, so there is no need for external equipment to conduct the necessary computation. Data acquisition was performed in a real environment using a unit installed into the surface of the pavement. A very large number of samples were collected containing measurements of various vehicle classes, which were applied for the training and the validation of the proposed algorithm. To explore the capabilities of magnetometers, nine defined vehicle classes were applied, which is much higher than in relevant methods. The classification is performed using three-layer feedforward artificial neural networks (ANN). Only time-domain analysis was performed on the waveforms using multiple novel feature extraction approaches. The applied time-domain features require low computation and memory resources, which enables easier implementation and real-time operation. Various combinations of used sensor axes were also examined to reduce the size of the classifier and to increase efficiency. The effect of the detection length, which is a widely used feature, but also speed-dependent, on the proposed system was also investigated to explore the suitability of the applied feature set. The results show that the highest achieved classification efficiencies on unknown samples are 74.67% with, and 73.73% without applying the detection length in the feature set.

Artificial Neural Network Assisted Multiobjective Optimization of Postharvest Blanching and Drying of Blueberries.

This study aimed to optimize the postharvest blanching and drying process of blueberries using high-humidity air impingement (HHAIB) and hot-air-assisted infrared (HAIR) heating. A novel pilot-scale hot-air-assisted carbon-fiber infrared (IR) blanching/drying system was developed. Fresh blueberries with an average diameter of 10~15 mm were first blanched with high-humidity air at 110 °C and 12 m/s velocity for different durations (30, 60, 90, and 120 s); subsequently, the preblanched blueberries were dried at different IR heating temperatures (50, 60, 70, 80, and 90 °C) and air velocities (0.01, 0.5, 1.5, and 2.5 m/s), following a factorial design. The drying time (DT), specific energy consumption (SEC), ascorbic acid content (VC), and rehydration capacity (RC) were determined as response variables. A three-layer feed-forward artificial neural network (ANN) model with a backpropagation algorithm was constructed to simulate the influence of blanching time, IR heating temperature, and air velocity on the four response variables by training on the experimental data. Objective functions for DT, SEC, VC, and RC that were developed by the ANN model were used for the simultaneous minimization of DT and SEC and maximization of VC and RC using a nondominated sorting genetic algorithm (NSGA II) to find the Pareto-optimal solutions. The optimal conditions were found to be 93 s of blanching, 89 °C IR heating, and a 1.2 m/s air velocity, which resulted in a drying time of 366.7 min, an SEC of 1.43 MJ/kg, a VC of 4.19 mg/100g, and an RC of 3.35. The predicted values from the ANN model agreed well with the experimental data under optimized conditions, with a low relative deviation value of 1.43-3.08%. The findings from this study provide guidance to improve the processing efficiency, product quality, and sustainability of blueberry postharvest processes. The ANN-assisted optimization approach developed in this study also sets a foundation for the smart control of processing systems of blueberries and similar commodities.

ANN AND ANALYTICAL SOLUTIONS TO RELATIVISTIC ISOTHERMAL GAS SPHERES

Relativistic isothermal gas spheres are a powerful tool to model many astronomical objects, like compact stars and clusters of galaxies. In the present paper, we introduce an artificial neural network (ANN) algorithm and Taylor series to model the relativistic gas spheres using Tolman-Oppenheimer-Volkoff differential equations (TOV). Comparing the analytical solutions with the numerical ones revealed good agreement with maximum relative errors of 10−3. The ANN algorithm implements a three-layer feed-forward neural network built using a back-propagation learning technique that is based on the gradient descent rule. We analyzed the massradius relations and the density profiles of the relativistic isothermal gas spheres against different relativistic parameters and compared the ANN solutions with the analytical ones. The comparison between the two solutions reflects the efficiency of using the ANN to solve TOV equations.

Ultrasound enhanced butyric acid-lauric acid designer lipid synthesis: Based on artificial neural network and changes in enzymatic structure

Ultrasound is a green technology for intensifying enzymatic reactions. In this study, an ultrasonic water bath with equipment parameters of 28 kHz, 1750.1 W/m2, 60% duty cycle was used to assist the synthesis of butyric acid-lauric acid designer lipid (BLDL), which was catalyzed by Lipozyme 435. A convincing three-layer feed-forward artificial neural network (ANN) model was established (R2 = 0.949, RMSE = 4.759, ADD = 7.329) to accurately predict the optimal parameters combination, which was described as 13.72 mL reaction volume, 15.49% enzyme loading, 0.253 substrate molar ratio (tributyrin/lauric acid), 56.58 °C reaction temperature and 120 min reaction time. The ultrasonic assistance increased actual butyric acid conversion rate by 11.38%, and also enhanced the consumption rate of tributyrin and lauric acid during the reaction. Meanwhile, the esterification activity of Lipozyme 435 was enhanced and its effectiveness up to 6 cycles. Structurally, ultrasound assistance significantly disrupted the secondary structure of the Lipozyme 435: reduced the content of α-helices, increased the content of β-sheet and β-turn. In addition, sonication caused an increase in crevice and micro-damage on the surface of the immobilized enzyme. In conclusion, low-intensity ultrasound at 28 kHz improved the synthesis efficiency of BLDL, which was scientifically predicted by ANN model, and the change of enzyme structure may be the vital reason for ultrasound enhanced reaction. However, the effect of ultrasound on immobilized enzymes’ activity needs to be further explored.

Quantum-inspired complex convolutional neural networks

Quantum-inspired neural network is one of the interesting researches at the junction of the two fields of quantum computing and deep learning. Several models of quantum-inspired neurons with real parameters have been proposed, which are mainly used for three-layer feedforward neural networks. In this work, we improve the quantum-inspired neurons by exploiting the complex-valued weights which have richer representational capacity and better non-linearity. We then extend the method of implementing the quantum-inspired neurons to the convolutional operations, and naturally draw the models of quantum-inspired convolutional neural networks (QICNNs) capable of processing high-dimensional data. Five specific structures of QICNNs are discussed which are different in the way of implementing the convolutional and fully connected layers. The performance of classification accuracy of the five QICNNs are tested on the MNIST and CIFAR-10 datasets. The results show that the QICNNs can perform better in classification accuracy on MNIST dataset than the classical CNN. More learning tasks that our QICNN can outperform the classical counterparts will be found.

Feedback Control Systems Stabilization Using a Bio-inspired Neural Network

The proportional–integral–derivative (PID) control systems, which have become a standard for technical and industrial applications, are the fundamental building blocks of classical and modern control systems. In this paper, a three-layer feed-forward neural network (NN) model trained to replicate the behavior of a PID controller is employed to stabilize control systems through a NN feedback controller. A novel bio-inspired weights-and-structure-determination (BIWASD) algorithm, which incorporates a metaheuristic optimization algorithm dubbed beetle antennae search (BAS), is used to train the NN model. More presicely, the BIWASD algorithm identifies the ideal weights and structure of the BIWASD-based NN (BIWASDNN) model utilizing a power sigmoid activation function while handling model fitting and validation. The results of three simulated trials on stabilizing feedback control systems validate and demonstrate the BIWASDNN model’s exceptional learning and prediction capabilities, while achieving similar or better performance than the corresponding PID controller. The BIWASDNN model is compared to three other high-performing NN models, and a MATLAB repository is accessible in public through GitHub to encourage and enhance this work.

Formation of the quadcopter flight path under overland monitoring using neuro-fuzzy modeling methods

Some territories where economic activity is carried out are characterized by the presence of mountainі and forests. To provide information support for the development of infrastructure and agriculture in these areas, in some cases, there is required overland monitoring with un-manned technologies. In this regard, an algorithm for the formation of a 3D trajectory of a quadcopter during overland piloting in a mountainous and wooded landscape is proposed, which implies autonomous maneuvering to overcome possible obstacles. As a basic model, it is pro-posed to use a Fuzzy Inference System with input characteristics in the form of linguistic varia-bles that reflect fuzzy sectors of space, within which the presence of obstacles and the distance to them are interpreted verbally, i.e., in the form of terms of corresponding input linguistic var-iables. Overcoming obstacles is supposed to be performed on the basis of fuzzy conclusions of the proposed system, formulated as terms of output linguistic variables which reflect changes in the angle of rotation in the horizontal plane, flight altitude, and traverse speed of the quadcop-ter. The paper analyzes the results of the model behavior for different scenarios of the terms of the input linguistic variables. For the operational formation of the quadcopter flight path, it is also proposed to use neural network modeling tools. Due to its ability to adapt to new condi-tions and requirements, the neural network model can become an important tool in ensuring the autonomous flight of a quadcopter under overland monitoring. For appropriate training of the three-layer feedforward neural network, a sufficiently large number of quadcopter behavior scenarios are used, which were generated by the Fuzzy Inference System relative to the over-coming of possible obstacles in five sectors of the survey.

A two-stage reliability allocation method for remanufactured machine tools integrating neural networks and remanufacturing coefficient

Reliability allocation is an important task that needs to be done in the design phase of machine tool remanufacturing to ensure that remanufactured machine tools meet the reliability target. However, unlike new machine tool products, remanufactured machine tools have high uncertainty and small samples, and traditional reliability methods are not suitable for remanufactured machine tools. This paper aims to propose an improved reliability allocation method for remanufactured machine tools integrating neural networks and remanufacturing coefficient. With the fault tree analysis (FTA) model constructed, the fault of remanufactured machine tools can be divided into three levels: system-level, subsystem-level, and part-level. The three-layer feedforward artificial neural network is adopted to allocate system reliability to subsystem-level. When reliability is allocated from subsystem-level to part-level, the remanufacturing comprehensive evaluation system and a remanufacturing coefficient that takes into account the characteristics of remanufactured components are introduced. Finally, the proposed method is illustrated in a case of reliability allocation for remanufactured NC gear hobbing machines. Moreover, the results show that the reliability target can be achieved and the growth of reliability can be guaranteed through the proposed method.

Dynamics and Control of a Magnetic Transducer Array Using Multi-Physics Models and Artificial Neural Networks.

A linear mechanical oscillator is non-linearly coupled with an electromagnet and its driving circuit through a magnetic field. The resulting non-linear dynamics are investigated using magnetic circuit approximations without major loss of accuracy and in the interest of brevity. Different computational approaches to simulate the setup in terms of dynamical system response and design parameters optimization are pursued. A current source operating in baseband without modulation directly feeds the electromagnet, which consists commonly of a solenoid and a horseshoe-shaped core. The electromagnet is then magnetically coupled to a mass made of soft magnetic material and attached to a spring with damping. The non-linear system is described by a linearized steady-space representation while is examined for controllability and observability. A controller using a pole placement approach is built to stabilize the element. Drawing upon the fact that coupling works both ways, enabling estimation of the mass position and velocity (state variables) by processing the induced voltage across the electromagnet, a state observer is constructed. Accurate and fast tracking of the state variables, along with the possibility of driving more than one module from the same source using modulation, proves the applicability of the electro-magneto-mechanical transducer for sensor applications. Next, a three-layer feed-forward artificial neural network (ANN) system equivalent was trained using the non-linear plant-linear controller-linear observer configuration. Simulations to investigate the robustness of the system with respect to different equilibrium points and input currents were carried out. The ANN proved robust with respect to position accuracy.

A Comparison of In-Sample and Out-of-Sample Model Selection Approaches for Artificial Neural Network (ANN) Daily Streamflow Simulation

Artificial Neural Networks (ANN) have been widely applied in hydrologic and water quality (H/WQ) modeling in the past three decades. Many studies have demonstrated an ANN’s capability to successfully estimate daily streamflow from meteorological data on the watershed level. One major challenge of ANN streamflow modeling is finding the optimal network structure with good generalization capability while ameliorating model overfitting. This study empirically examines two types of model selection approaches for simulating streamflow time series: the out-of-sample approach using blocked cross-validation (BlockedCV) and an in-sample approach that is based on Akaike’s information criterion (AIC) and Bayesian information criterion (BIC). A three-layer feed-forward neural network using a back-propagation algorithm is utilized to create the streamflow models in this study. The rainfall–streamflow relationship of two adjacent, small watersheds in the San Antonio region in south-central Texas are modeled on a daily time scale. The model selection results of the two approaches are compared, and some commonly used performance measures (PMs) are generated on the stand-alone testing datasets to evaluate the models selected by the two approaches. This study finds that, in general, the out-of-sample and in-sample approaches do not converge to the same model selection results, with AIC and BIC selecting simpler models than BlockedCV. The ANNs were found to have good performance in both study watersheds, with BlockedCV selected models having a Nash–Sutcliffe coefficient of efficiency (NSE) of 0.581 and 0.658, and AIC/BIC selected models having a poorer NSE of 0.574 and 0.310, for the two study watersheds. Overall, out-of-sample BlockedCV selected models with better predictive ability and is preferable to model streamflow time series.

Evaluation of Fischer-Tropsch synthesis to light olefins over Co- and Fe-based catalysts using artificial neural network

Currently, energy transition due to fossil fuel negative side effects is taking place. This transition impacts the chemical industry based on light olefins reactions. Plastics, detergents, polymers, and others are mostly produced from such hydrocarbons, which are mainly originated from oil-based and highly energy-consuming processes. Fischer-Tropsch Synthesis (FTS) is a strategic technology capable to transform a given carbon source, including natural gas and biomass, into high added-value hydrocarbons. It is affected by several conditions, such as catalyst design and operating conditions and its feasibility requires a good selection of relevant process variables to optimize light olefins yield. In this work, Machine Learning models were used to predict adequate reaction conditions from the catalytic literature data. Three-layer feedforward neural networks were adjusted using a careful selection of operating conditions and catalyst composition as inputs and carbon monoxide conversion, light olefins selectivity, and carbon dioxide yield as outputs. The results indicate neural network prediction efficacy for FTS most relevant variables, such as temperature and catalyst composition. This work presents the novelty of including more variables in the model compared to recent similar studies, such as catalyst support, active phase, and promoters as inputs; and light olefins selectivity and CO2 yield as outputs. Overall, Fe-based catalyst (standard Fe (1.6 wt%) K/TiO2) presented the highest light olefins selectivity and yield at optimal conditions (T = 500 °C and 20 wt% of active phase), despite showing the highest emission of carbon dioxide.

Intermittent evolution routes to the periodic or the chaotic orbits in Rulkov map.

This paper concerns the intermittent evolution routes to the asymptotic regimes in the Rulkov map. That is, the windows with transient approximate periodic and transient chaotic behaviors occur alternatively before the system reaches the periodic or the chaotic orbits. Meanwhile, the evolution routes to chaotic orbits can be classified into different types according to the windows before reaching asymptotic chaotic states. In addition, the initial values can be regarded as a key factor affecting the asymptotic behaviors and the evolution routes. The effects of the initial values are given by parameter planes, bifurcation diagrams, and waveforms. In order to investigate whether the intermittent evolution routes can be learned by machine learning, some experiments are given to understanding the differences between the trajectories of the Rulkov map generated by the numerical simulations and predicted by the neural networks. These results show that there is about 60% accuracy rate of successfully predicting both the evolution routes and the asymptotic period-3 orbits using a three-layer feedforward neural network, while the bifurcation diagrams can be reconstructed using reservoir computing except a few parameter conditions.

Application of Artificial Neural Networks to Predict the Intrinsic Solubility of Drug-Like Molecules.

Machine learning (ML) approaches are receiving increasing attention from pharmaceutical companies and regulatory agencies, given their ability to mine knowledge from available data. In drug discovery, for example, they are employed in quantitative structure–property relationship (QSPR) models to predict biological properties from the chemical structure of a drug molecule. In this paper, following the Second Solubility Challenge (SC-2), a QSPR model based on artificial neural networks (ANNs) was built to predict the intrinsic solubility (logS0) of the 100-compound low-variance tight set and the 32-compound high-variance loose set provided by SC-2 as test datasets. First, a training dataset of 270 drug-like molecules with logS0 value experimentally determined was gathered from the literature. Then, a standard three-layer feed-forward neural network was defined by using 10 ChemGPS physico-chemical descriptors as input features. The developed ANN showed adequate predictive performances on both of the SC-2 test datasets. Benefits and limitations of ML approaches have been highlighted and discussed, starting from this case-study. The main findings confirmed that ML approaches are an attractive and promising tool to predict logS0; however, many aspects, such as data quality, molecular descriptor computation and selection, and assessment of applicability domain, are crucial but often neglected, and should be carefully considered to improve predictions based on ML.