Can it?
Last updated
Last updated
Yes, Predictor can predict minimal values such as Ultimate Tensile Strength (UTS) and Yield Strength (YS) for materials under specific conditions. These predictions are typically presented as minimum values, indicated by the ">" symbol on the prediction page. For instance and as shown in the screenshot below, when UTS is predicted for a material at Room Temperature (RT), it may be displayed as "> 400.161 MPa," indicating that the minimum tensile strength is above this value under the given conditions.
Yes, Predictor's so called Synthetic models can predict the most likely ranges for mechanical properties such as UTS and YS even without detailed knowledge of the material's heat treatment or processing history. These models use only the chemical composition or material designation to provide property estimates at Room Temperature (RT), making it particularly useful when processing details are unavailable.
Yes, the reliability of predictions from each model can be assessed using performance metrics values obtained on model testing.
To estimate the performance and applicability of a machine learning model, various numerical metrics must be used to evaluate how well the model performs. These metrics help quantify the model's prediction accuracy and provide insight into its strengths, weaknesses, and utility for specific tasks.
Predictor uses several regression metrics, and they each provide a different perspective on model performance:
MAPE, MAE, and RMSE are similar, whereby RMSE is more sensitive to high residuals, while MAPE is dimensionless, i.e. it is expressed in %,
Because of being dimensionless, as well as r and R², MAPE is convenient for comparing the performance of different models, even if they predict different properties,
RMSE and MAE provide a more meaningful interpretation because they are measured in the units of the predicted property,
Also, the MAPE, RMSE, and MAE can be used to calculate the confidence interval,
Finally, a metric that provides a share of predictions within relative error allows flexible tolerance evaluation (error margin 5%, 10%, or 20%) for different precision requirements, making it easier to understand and communicate the model’s effectiveness.
Yes, the reliability of predictions is assessed through two key methods: Model Applicability and Point-Based Confidence Intervals.
Model Applicability: This determines how well the model applies to the selected material and conditions. There are three levels of applicability, which are highlighted in the system with different colors:
Fully applicable - All inputs are within the model's testing domain.
Applicable - Values of all inputs are within the model scope (training domain), but one or more inputs are outside of the testing domain.
Extrapolation - Values of one or more inputs are outside the model's scope (training domain), meaning the prediction involves extrapolation beyond the tested range.
Point-Based Confidence Intervals: Confidence intervals offer statistical measures of prediction accuracy, ranging from 80% to 99.9%. These intervals are calculated by dividing the dataset into regions using the Mahalanobis distance, a method of measuring the distance between a data point and a distribution. Each region has a specific confidence interval based on testing results. When a prediction is made, the system determines the Mahalanobis distance for the input and assigns the corresponding confidence interval to the predicted value.
No, the application does not automatically filter or suggest suitable heat treatments or operating temperatures for specific materials. While the model displays the temperature scope (e.g., "The temperature range is from -273°C to 497°C"), users must understand that not all materials in a given group can withstand the entire temperature range. Therefore, it is up to the user to assess the material's properties and limitations when making predictions.
No, predicting a chemical composition based solely on a set of target values for material properties is a complex challenge due to the multiplicity of solutions. This challenge arises because many different chemical compositions can exhibit similar physical or mechanical properties, meaning there isn't a single, unique composition that corresponds to a particular set of target properties. As a result, numerous potential compositions could satisfy the same property requirements, leading to an almost infinite range of possible solutions.
In contrast, predicting material properties from a known chemical composition and material condition (e.g., heat treatment or form) is more straightforward. When these inputs are provided, the system can often offer a narrowed-down or unique set of predictions for properties, as the problem becomes more deterministic. This stands in contrast to trying to reverse-engineer the chemical composition based purely on desired properties, which introduces significant ambiguity and complexity. Therefore, Predictor does not currently support the process of deducing a chemical composition from target properties.
In the broader field of materials science, this issue illustrates how the relationship between properties and composition is multifaceted and indirect. For instance, when target properties are input into a system like Predictor, it becomes evident that different compositions can lead to similar results. However, when starting with a known chemical composition and specific conditions, the prediction of material properties is much more precise and deterministic, reflecting a more direct relationship between input and output.
Yes, the integration of Predictor with the Diagram Comparison tool provides significant advantages when selecting the optimal heat treatment for a material. This feature allows users to predict material properties under various heat treatment scenarios and compare them side-by-side using the Diagram Comparison tool. This functionality not only facilitates a comprehensive evaluation of how different heat treatments affect material properties but also allows for the juxtaposition of these predictions against actual property curves from the material records within Total Materia Horizon.
Users can:
Run successive predictions of how different heat treatments will impact material properties.
Add these predictions to the Diagram Comparison tool for direct comparison with actual property curves from the Total Materia Horizon database.
Evaluate how each heat treatment affects the material’s mechanical properties, such as tensile strength or hardness, under different conditions.
This combination of predictive analysis and real-world data allows for a comprehensive evaluation of the impact of various heat treatments, helping users make informed decisions. By comparing predictions against existing material data, engineers can determine the best possible heat treatment for specific materials to ensure optimal performance in their applications. This functionality is available as part of the user’s subscription to the Total Materia platform, enhancing its utility for material selection and performance optimization.
