Machine Learning Experiments Research Articles

Exploring the CO2 conversion activated by the dielectric barrier discharge plasma assisted with photocatalyst via machine learning

DBD plasma assisted with halide perovskite photocatalysts is very attractive for the conversion of CO2 into high value-added products under mild conditions. However, it is still challenging to further effectively improve the corresponding CO2 conversion efficiency due to the consumption of time, materials, equipments via traditional experimental approach. In this paper, we simulate the CO2 conversion process under the DBD plasma assisted with different photocatalysts by machine learning. The process conditions and material parameters were selected as the feature variables of the machine learning model, and all the features were screened by combining Pearson's correlation coefficient and MIR ranking. The regression algorithms were evaluated using K-fold cross-validation combined with the coefficient of determination (R2), while SVR and BPANN with the best prediction performance for CO2 conversion ratio and energy efficiency were selected to establish the prediction models. In order to increase the accuracy of predictions, the hyperparameters of these two base models were improved using genetic algorithm, particle swarm optimization and Bayesian optimization. The R2 of the GA-SVR-CO2 and Bayesian-BPANN-EE reached 0.8960 and 0.9679 for the testing sets, respectively, and their practical applications were successfully verified. In addition, the effects of feature parameters on the conversion efficiency were revealed by SHAP. SEI scatter plots and Spearman correlation coefficient have been used for correlation analysis to better explore the correlation between machine learning models and experiments. This work brings new insights into the contributions of multiple parameters in the plasma system assisted with photocatalyst for efficient conversion of CO2.

Acoustic encoding of vocally expressed confidence and doubt in Chinese bidialectics.

Language communicators use acoustic-phonetic cues to convey a variety of social information in the spoken language, and the learning of a second language affects speech production in a social setting. It remains unclear how speaking different dialects could affect the acoustic metrics underlying the intended communicative meanings. Nine Chinese Bayannur-Mandarin bidialectics produced single-digit numbers in statements of both Standard Mandarin and the Bayannur dialect with different levels of intended confidence. Fifteen listeners judged the intention presence and confidence level. Prosodically unmarked and marked stimuli exhibited significant differences in perceived intention. A higher intended level was perceived as more confident. The acoustic analysis revealed the segmental (third and fourth formants, center of gravity), suprasegmental (mean fundamental frequency, fundamental frequency range, duration), and source features (harmonic to noise ratio, cepstral peak prominence) can distinguish between confident and doubtful expressions. Most features also distinguished between dialect and Mandarin productions. Interactions on fourth formant and mean fundamental frequency suggested that speakers made greater use of acoustic parameters to encode confidence and doubt in the Bayannur dialect than in Mandarin. In machine learning experiments, the above-chance-level overall classification rates for confidence and doubt and the in-group advantage supported the dialect theory.

Stance detection in Arabic with a multi-dialectal cross-domain stance corpus

We present a cross-domain and multi-dialectal stance corpus for Arabic, covering the major dialect groups and four Arab regions. This research provides an important language resource for automating the task of stance detection in Dialectal Arabic while carefully considering the subtle differences in stance expression across various dialects. More than 4500 sentences in our corpus have been carefully annotated according to their stance with regard to a certain subject. We gathered sentences associated with two controversial topics for every region and we had at least two annotators annotate each sentence to indicate if the author is supporting, opposing, or neutral to the sentence’s topic. Our corpus shows high balance between dialect and stance. About half of the sentences in each region are written in Modern Standard Arabic, while the other half are written in the specific dialect of that region. To evaluate our corpus, we performed a number of machine-learning experiments for the stance detection task. The best performance was achieved by AraBERT with an accuracy and an F1-score of 0.82. Furthermore, we trained and tested this model on the most similar state-of-the-art stance dataset, “MAWQIF”. The comparison results demonstrate how crucial it is to maintain balance among the three stance classes in our dataset. In particular, the model scored better when using our stance corpus than when using the MAWQIF dataset especially for the “Neutral” stance class. Using our best performing model, we developed a Web-based demonstrator for stance detection in dialectal Arabic and we show its effectiveness in analyzing stance in the context of two real-world scenarios: product boycott in the Arab world and customer reviews of a soft drink company.

Bridging length scales in hard materials with ultra-small angle X-ray scattering - a critical review.

Owing to their exceptional properties, hard materials such as advanced ceramics, metals and composites have enormous economic and societal value, with applications across numerous industries. Understanding their microstructural characteristics is crucial for enhancing their performance, materials development and unleashing their potential for future innovative applications. However, their microstructures are unambiguously hierarchical and typically span several length scales, from sub-ångstrom to micrometres, posing demanding challenges for their characterization, especially for in situ characterization which is critical to understanding the kinetic processes controlling microstructure formation. This review provides a comprehensive description of the rapidly developing technique of ultra-small angle X-ray scattering (USAXS), a nondestructive method for probing the nano-to-micrometre scale features of hard materials. USAXS and its complementary techniques, when developed for and applied to hard materials, offer valuable insights into their porosity, grain size, phase composition and inhomogeneities. We discuss the fundamental principles, instrumentation, advantages, challenges and global status of USAXS for hard materials. Using selected examples, we demonstrate the potential of this technique for unveiling the microstructural characteristics of hard materials and its relevance to advanced materials development and manufacturing process optimization. We also provide our perspective on the opportunities and challenges for the continued development of USAXS, including multimodal characterization, coherent scattering, time-resolved studies, machine learning and autonomous experiments. Our goal is to stimulate further implementation and exploration of USAXS techniques and inspire their broader adoption across various domains of hard materials science, thereby driving the field toward discoveries and further developments.

Unraveling pathogenesis, biomarkers and potential therapeutic agents for endometriosis associated with disulfidptosis based on bioinformatics analysis, machine learning and experiment validation

BackgroundEndometriosis (EMs) is an enigmatic disease of yet-unknown pathogenesis. Disulfidptosis, a novel identified form of programmed cell death resulting from disulfide stress, stands a chance of treating diverse ailments. However, the potential roles of disulfidptosis-related genes (DRGs) in EMs remain elusive. This study aims to thoroughly explore the key disulfidptosis genes involved in EMs, and probe novel diagnostic markers and candidate therapeutic compounds from the aspect of disulfidptosis based on bioinformatics analysis, machine learning, and animal experiments.ResultsEnrichment analysis on key module genes and differentially expressed genes (DEGs) of eutopic and ectopic endometrial tissues in EMs suggested that EMs was closely related to disulfidptosis. And then, we obtained 20 and 16 disulfidptosis-related DEGs in eutopic and ectopic endometrial tissue, respectively. The protein-protein interaction (PPI) network revealed complex interactions between genes, and screened nine and ten hub genes in eutopic and ectopic endometrial tissue, respectively. Furthermore, immune infiltration analysis uncovered distinct differences in the immunocyte, human leukocyte antigen (HLA) gene set, and immune checkpoints in the eutopic and ectopic endometrial tissues when compared with health control. Besides, the hub genes mentioned above showed a close correlation with the immune microenvironment of EMs. Furthermore, four machine learning algorithms were applied to screen signature genes in eutopic and ectopic endometrial tissue, including the binary logistic regression (BLR), the least absolute shrinkage and selection operator (LASSO), the support vector machine-recursive feature elimination (SVM-RFE), and the extreme gradient boosting (XGBoost). Model training and hyperparameter tuning were implemented on 80% of the data using a ten-fold cross-validation method, and tested in the testing sets which determined the excellent diagnostic performance of these models by six indicators (Sensitivity, Specificity, Positive Predictive Value, Negative Predictive Value, Accuracy, and Area Under Curve). And seven eutopic signature genes (ACTB, GYS1, IQGAP1, MYH10, NUBPL, SLC7A11, TLN1) and five ectopic signature genes (CAPZB, CD2AP, MYH10, OXSM, PDLIM1) were finally identified based on machine learning. The independent validation dataset also showed high accuracy of the signature genes (IQGAP1, SLC7A11, CD2AP, MYH10, PDLIM1) in predicting EMs. Moreover, we screened 12 specific compounds for EMs based on ectopic signature genes and the pharmacological impact of tretinoin on signature genes was further verified in the ectopic lesion in the EMs murine model.ConclusionThis study verified a close association between disulfidptosis and EMs based on bioinformatics analysis, machine learning, and animal experiments. Further investigation on the biological mechanism of disulfidptosis in EMs is anticipated to yield novel advancements for searching for potential diagnostic biomarkers and revolutionary therapeutic approaches in EMs.

From Characterization to Discovery: Artificial Intelligence, Machine Learning and High-Throughput Experiments for Heterogeneous Catalyst Design

From Characterization to Discovery: Artificial Intelligence, Machine Learning and High-Throughput Experiments for Heterogeneous Catalyst Design

Chaos, overfitting and equilibrium: To what extent can machine learning beat the financial market?

In this paper, we applied 10 technical analysis indicators to predict stock price movement directions using support vector machines, investigating the effects of hyperparameter variations on the out-of-sample classification performance and the profitability of the resulting trading strategies. We collected daily data between January 1st, 2018, and March 31st, 2023 for the 30 firms that compose the Dow Jones Industrial Average (DJIA). Our results indicated that the out-of-sample accuracy converged to 50%, while a small percentage (13.63% for the pre-COVID period and 23.16% for the post-COVID period) of the hyperparameter combinations yielded gains above the buy-and-hold strategy; on the other hand, no clear patterns about the best-performing hyperparameter combinations emerged, as the behavior of the out-of-sample performance was found to exhibit high sensitive dependence to the hyperparameters settings in comparison to its in-sample counterpart. The outcomes of our empirical analysis are consistent with both classic results in the finance literature (such as the Efficient Market Hypothesis) and empirical setbacks commonly seen in machine learning experiments, notably the occurrence of overfitting under the incorporation of high-dimensional non-linear interactions.

The origin and function of external representations

External representations ( ERs) are objects or performances in the world whose proper function is to communicate about other things in the world. Why and how did we make them, and what do they give us? We outline a simple framework for thinking about ERs grounded in modern machine learning. We propose a minimal set of neural mechanisms needed for open-ended ER production. Our constructivist enactive view contrasts with nativist views which propose specialist neural modules requiring symbolic internal representations. We propose a plausible set of (biological and cultural) evolutionary steps to full Gricean symbolic communication via a set of increasingly complex enactive algorithms for ER production. A pragmatic space of games is defined, which includes not only fully cooperative language games but also science, art, and evolved signal manipulation games. This space is defined by the complexity of learning needed by sender and receiver. We propose that one important step towards open-ended ER use was selection for bush reading, which like mind-reading is an inferential process requiring complex contextual and syntactic understanding of cues about events displaced in space and time. Bush reading pre-adapted receivers to be receptive, competent, and perspicacious interpreters of later intentionally produced signals about hidden topics such as felt mental states. This paved the way for minimally Gricean communication, which subsequently could be bootstrapped into explicit theories of mind, folk psychology narratives, and symbolic language in general. Recent findings in cognitive archaeology are integrated within the framework, and new experiments in machine learning suggested.

Mechanism of Action of Curcumin for Rheumatoid Arthritis Based on Machine Learning, Molecular Dynamics and Cellular Experiments

<i>Objective</i>: Curcumin has been widely used in rheumatoid arthritis (RA) treatment and the previous study also proved its effectiveness. However, the pharmacological mechanism is still not clear. The current study intends to discuss the potential mechanism of action of curcumin in RA treatment through machine learning, network pharmacology, molecular dynamics and cellular experiments.<i>Methods</i>: RA-related microarray data were obtained from three GEO datasets: GSE55235, GSE55457 and GSE77298. Machine learning methods including XGBOOST, LASSO and SVM were adopted to screen out potential targets of RA pathogenesis. Online tools SwissTargetPrediction and Similarity ensemble approach were visited to predict potential targets of action of curcumin. The key target was identified via a Venn diagram and processed for molecular docking and molecular dynamics simulation with curcumin. Fibroblast-like synoviocytes (RA-FLSs) were selected to study the effect of curcumin at different concentrations (20, 40 and 80 μmol/L) on cell proliferation and apoptosis using MTT and flow cytometry assays. In addition, Western blot was used to examine the protein level.<i>Results</i>: Arachidonate 5-Lipoxygenase (ALOX5) was identified as a key target of RA following bioinformatics prediction. Results of molecular docking and molecular dynamics simulation demonstrated the tight binding between curcumin and ALOX5 with stable function. RA-FLSs intervened with different concentrations of curcumin (20, 40 and 80 mol/L) exhibited decreased potential in proliferation while increased apoptosis, which were in a dose-dependent manner. Additionally, with the increase of curcumin concentration, the protein level of ALOX5 gradually decreased. <i>Conclusion</i>: Curcumin may exert its therapeutic effects in RA treatment via down-regulating the expression of ALOX5.

Pharmaceutical consumption, economic growth and life expectancy in the OECD: the application of a new causal direction from dependency algorithm and a DeepNet process

PurposeThe objective of this study is to reevaluate the correlation among pharmaceutical consumption, per capita income, and life expectancy across different age groups (at birth, middle age, and advanced age) within the OECD countries between 1998 and 2018.Design/methodology/approachWe employ a two-step methodology, utilizing two independent approaches. Firstly, we con-duct the Dumitrescu-Hurlin pairwise panel causality test, followed by Machine Learning (ML) experiments employing the Causal Direction from Dependency (D2C) Prediction algorithm and a DeepNet process, thought to deliver robust inferences with respect to the nature, sign, direction, and significance of the causal relationships revealed in the econometric procedure.FindingsOur findings reveal a two-way positive bidirectional causal relationship between GDP and total pharmaceutical sales per capita. This contradicts the conventional notion that health expenditures decrease with economic development due to general health improvements. Furthermore, we observe that GDP per capita positively correlates with life expectancy at birth, 40, and 60, consistently generating positive and statistically significant predictive values. Nonetheless, the value generated by the input life expectancy at 60 on the target income per capita is negative (−61.89%), shedding light on the asymmetric and nonlinear nature of this nexus. Finally, pharmaceutical sales per capita improve life expectancy at birth, 40, and 60, with higher magnitudes compared to those generated by the income input.Practical implicationsThese results offer valuable insights into the intricate dynamics between economic development, pharmaceutical consumption, and life expectancy, providing important implications for health policy formulation.Originality/valueVery few studies shed light on the nature and the direction of the causal relationships that operate among these indicators. Exiting from the standard procedures of cross-country regressions and panel estimations, the present manuscript strives to promote the relevance of using causality tests and Machine Learning (ML) methods on this topic. Therefore, this paper seeks to contribute to the literature in three important ways. First, this is the first study analyzing the long-run interactions among pharmaceutical consumption, per capita income, and life expectancy for the Organization for Economic Co-operation and Development (OECD) area. Second, this research contrasts with previous ones as it employs a complete causality testing framework able to depict causality flows among multiple variables (Dumitrescu-Hurlin causality tests). Third, this study displays a last competitive edge as the panel data procedures are complemented with an advanced data testing method derived from AI. Indeed, using an ML experiment (i.e. Causal Direction from Dependency, D2C and algorithm) it is believed to deliver robust inferences regarding the nature and the direction of the causality. All in all, the present paper is believed to represent a fruitful methodological research orientation. Coupled with accurate data, this seeks to complement the literature with novel evidence and inclusive knowledge on this topic. Finally, to bring accurate results, data cover the most recent and available period for 22 OECD countries: from 1998 to 2018.

Exploring liquid-liquid phase separation-related diagnostic biomarkers in osteoarthritis based on machine learning algorithms and experiment

BackgroundOsteoarthritis (OA) is a prevalent joint disorder characterized by cartilage degeneration and joint inflammation. Liquid-liquid phase separation (LLPS), a biophysical process involved in cellular organization, has recently gained attention in OA research. However, the relationship between LLPS and OA remains poorly understood. MethodsWe analyzed gene expression data from the GSE48556 dataset to identify LLPS-related genes associated with OA. Differential expression analysis, enrichment analyses, and machine learning algorithms were employed to explore the functional significance of LLPS-related genes in OA and then construct a diagnostic model for OA. In addition, IL-1β as a pro-inflammatory factor to establish an in vitro OA model, and the protein expression levels of OA biomarkers were detected by western blot. ResultsA total of 145 LLPS-related genes were screened in OA samples. Enrichment analyses revealed these genes were mainly enriched in mRNA metabolic processes, cytoplasmic granules, and insulin resistance. Four characteristic genes for OA were selected by using machine learning algorithms, including ADRB2, H3F3B, GNL3L, and PELO. These genes showed satisfactory diagnostic values. Furthermore, there were association between these biomarkers and immune cells, including T cells CD8 and monocytes. In vitro experiments showed that IL-1β stimulation significantly inhibited the cell viability of chondrocytes and enhanced the levels of pro-inflammatory factors, that could mimic the inflammatory state of OA. The expression levels of GNL3L and H3F3B proteins in IL-1β group were obviously lower than those in control group, while levels of ADRB2 and PELO were higher, which was consistent with the results of bioinformatics analysis. ConclusionOur study identifies LLPS-related genes as potential diagnostic biomarkers for OA. These findings provide insights into the molecular mechanisms underlying OA pathogenesis and offer opportunities for the development of novel therapeutic strategies.

Optimizing clinico-genomic disease prediction across ancestries: a machine learning strategy with Pareto improvement

BackgroundAccurate prediction of an individual’s predisposition to diseases is vital for preventive medicine and early intervention. Various statistical and machine learning models have been developed for disease prediction using clinico-genomic data. However, the accuracy of clinico-genomic prediction of diseases may vary significantly across ancestry groups due to their unequal representation in clinical genomic datasets.MethodsWe introduced a deep transfer learning approach to improve the performance of clinico-genomic prediction models for data-disadvantaged ancestry groups. We conducted machine learning experiments on multi-ancestral genomic datasets of lung cancer, prostate cancer, and Alzheimer’s disease, as well as on synthetic datasets with built-in data inequality and distribution shifts across ancestry groups.ResultsDeep transfer learning significantly improved disease prediction accuracy for data-disadvantaged populations in our multi-ancestral machine learning experiments. In contrast, transfer learning based on linear frameworks did not achieve comparable improvements for these data-disadvantaged populations.ConclusionsThis study shows that deep transfer learning can enhance fairness in multi-ancestral machine learning by improving prediction accuracy for data-disadvantaged populations without compromising prediction accuracy for other populations, thus providing a Pareto improvement towards equitable clinico-genomic prediction of diseases.

Unlocking the Future: Privacy-Preserving ML Experimentation

Abstract: Experiments with machine learning (ML) have become a key source of new ideas in many fields. However, growing worries about data privacy have made it clear that we need ML testing methods that protect privacy. There are new technologies in this piece that let you play around with machine learning without putting your data at risk. Differential privacy, secure multiparty computation (SMPC), homomorphic encryption, federated learning, trusted execution environments (TEEs), making fake data, and using temporary and nameless IDs are some of these technologies. By using these privacy-protecting solutions, businesses can utilize the full potential of machine learning experiments while protecting individuals' privacy rights and staying in line with strict rules.

Machine learning experiment management tools: a mixed-methods empirical study

Machine Learning (ML) experiment management tools support ML practitioners and software engineers when building intelligent software systems. By managing large numbers of ML experiments comprising many different ML assets, they not only facilitate engineering ML models and ML-enabled systems, but also managing their evolution—for instance, tracing system behavior to concrete experiments when the model performance drifts. However, while ML experiment management tools have become increasingly popular, little is known about their effectiveness in practice, as well as their actual benefits and challenges. We present a mixed-methods empirical study of experiment management tools and the support they provide to users. First, our survey of 81 ML practitioners sought to determine the benefits and challenges of ML experiment management and of the existing tool landscape. Second, a controlled experiment with 15 student developers investigated the effectiveness of ML experiment management tools. We learned that 70% of our survey respondents perform ML experiments using specialized tools, while out of those who do not use such tools, 52% are unaware of experiment management tools or of their benefits. The controlled experiment showed that experiment management tools offer valuable support to users to systematically track and retrieve ML assets. Using ML experiment management tools reduced error rates and increased completion rates. By presenting a user’s perspective on experiment management tools, and the first controlled experiment in this area, we hope that our results foster the adoption of these tools in practice, as well as they direct tool builders and researchers to improve the tool landscape overall.

Secondary organic aerosols derived from intermediate-volatility n-alkanes adopt low-viscous phase state

Abstract. Secondary organic aerosol (SOA) derived from n-alkanes, as emitted from vehicles and volatile chemical products, is a major component of anthropogenic particulate matter, yet the chemical composition and phase state are poorly understood and thus poorly constrained in aerosol models. Here we provide a comprehensive analysis of n-alkane SOA by explicit gas-phase chemistry modeling, machine learning, and laboratory experiments to show that n-alkane SOA adopts low-viscous semi-solid or liquid states. Our study underlines the complex interplay of molecular composition and SOA viscosity: n-alkane SOA with a higher carbon number mostly consists of less functionalized first-generation products with lower viscosity, while the SOA with a lower carbon number contains more functionalized multigenerational products with higher viscosity. This study opens up a new avenue for analysis of SOA processes, and the results indicate few kinetic limitations of mass accommodation in SOA formation, supporting the application of equilibrium partitioning for simulating n-alkane SOA formation in large-scale atmospheric models.

Dementia Detection with Phonetic and Phonological Features

In this paper the ADDReSS challenge dataset was used for training and testing a binary classifier designed to diagnose AD. This dataset consists of transcripts of descriptions of the Cookie Theft picture, produced by 54 subjects in the training part and 24 subjects in the test part. Two machine learning experiments were conducted on the task of classifying transcribed speech samples with text samples that were produced by people with AD from those produced by normal subjects. The first experiment showed that, among all the subtypes of phonetic and phonological features covered in this paper, vowels provided the best classification performance. The second experiment that used four feature selection techniques showed with the adopted phonetic and phonological features provided about 0.87 F1 score, that is close to the best performance reported in the address challenge by systems using multiple linguistic levels and machine learning techniques. This result confirms the importance of the covered features as indicators of dementia.

Naturalistic Digital Behavior Predicts Cognitive Abilities

Individuals are known to differ in cognitive abilities, affecting their behavior and information processing in digital environments. However, we have a limited understanding of which behaviors are affected, how, and whether some features extracted from digital behavior can predict cognitive abilities. Consequently, researchers may miss opportunities to design and support individuals with personalized experiences and detect those who may benefit from additional interventions. To characterize digital behaviors, we collected 24/7 screen recordings, input behavior, and operating system data from the laptops of 20 adults for two weeks. We use cognitive test results from the same individuals to characterize their cognitive abilities: psychomotor speed, processing speed, selective attention, working memory, and fluid intelligence. Our results from regression analysis, path modeling, and machine learning experiments show that cognitive abilities are associated with differences in digital behavior and that naturalistic behavioral data can predict the cognitive abilities of individuals with small error rates. Our findings suggest naturalistic interaction data as a novel source for modeling cognitive differences.

A Critical Review of the Modelling Tools for the Reactive Transport of Organic Contaminants

The pollution of groundwater and soil by hydrocarbons is a significant and growing global problem. Efforts to mitigate and minimise pollution risks are often based on modelling. Modelling-based solutions for prediction and control play a critical role in preserving dwindling water resources and facilitating remediation. The objectives of this article are to: (i) to provide a concise overview of the mechanisms that influence the migration of hydrocarbons in groundwater and to improve the understanding of the processes that affect contamination levels, (ii) to compile the most commonly used models to simulate the migration and fate of hydrocarbons in the subsurface; and (iii) to evaluate these solutions in terms of their functionality, limitations, and requirements. The aim of this article is to enable potential users to make an informed decision regarding the modelling approaches (deterministic, stochastic, and hybrid) and to match their expectations with the characteristics of the models. The review of 11 1D screening models, 18 deterministic models, 7 stochastic tools, and machine learning experiments aimed at modelling hydrocarbon migration in the subsurface should provide a solid basis for understanding the capabilities of each method and their potential applications.

Predicting root numbers with neural networks

In this paper, we report on two machine learning experiments in search of statistical relationships between Dirichlet coefficients and root numbers or analytic ranks of certain low-degree [Formula: see text]-functions. The first experiment is to construct interpretable models based on murmurations, a recently discovered correlation between Dirichlet coefficients and root numbers. We show experimentally that these models achieve high accuracy by learning a combination of Mestre–Nagao-type heuristics and murmurations, noting that the relative importance of these features varies with degree. The second experiment is to search for a low-complexity statistic of Dirichlet coefficients that can be used to predict root numbers in polynomial time. We give experimental evidence and provide heuristics that suggest this cannot be done with standard machine learning techniques.

Analysis of learning curves in predictive modeling using exponential curve fitting with an asymptotic approach.

The existence of large volumes of data has considerably alleviated concerns regarding the availability of sufficient data instances for machine learning experiments. Nevertheless, in certain contexts, addressing limited data availability may demand distinct strategies and efforts. Analyzing COVID-19 predictions at pandemic beginning emerged a question: how much data is needed to make reliable predictions? When does the volume of data provide a better understanding of the disease's evolution and, in turn, offer reliable forecasts? Given these questions, the objective of this study is to analyze learning curves obtained from predicting the incidence of COVID-19 in Brazilian States using ARIMA models with limited available data. To fulfill the objective, a retrospective exploration of COVID-19 incidence across the Brazilian States was performed. After the data acquisition and modeling, the model errors were assessed by employing a learning curve analysis. The asymptotic exponential curve fitting enabled the evaluation of the errors in different points, reflecting the increased available data over time. For a comprehensive understanding of the results at distinct stages of the time evolution, the average derivative of the curves and the equilibrium points were calculated, aimed to identify the convergence of the ARIMA models to a stable pattern. We observed differences in average derivatives and equilibrium values among the multiple samples. While both metrics ultimately confirmed the convergence to stability, the equilibrium points were more sensitive to changes in the models' accuracy and provided a better indication of the learning progress. The proposed method for constructing learning curves enabled consistent monitoring of prediction results, providing evidence-based understandings required for informed decision-making.