A New Era for solder fatigue and electronic reliability analysis

PEARL_Logo.jpg

PEARL-AI

Prediction of Electronic Advanced Reliability and Life using Artificial Intelligence

 

Electronic Reliability Background

Many industries develop and deploy high-reliability electronic assets, like satellites or payloads, to accomplish specific missions with high probability of success.  Just about every piece of electronic hardware out in the world has undergone an electronic reliability analysis, or ERA.  This analysis is performed to estimate the expected usable life of electronic hardware within its operational environment.  Naturally, systems which are more difficult to repair or replace are subject to stricter reliability requirements because of the custom nature of the hardware and the cost and complexity of actually getting hands-on maintenance access.  Hardware design and life requirements for high-reliability industries can differ substantially when compared to the automotive industry, for example, where the acceptable probability of hardware failure may be one-percent (or less) over its operational life.  These strict requirements are a result of the relative access difficulty, finite or nonexistent availability of replacement hardware, and specialized operations knowledge among limited capable maintenance facilities or personnel. 

 

The broader scope of ERA encompasses a wide variety of analysis techniques and empirical models based on a products Lifecycle Environmental Profile, or LCEP, but the primary three types of environments are shock, vibration, and thermal cycling.  Each of these environments can cause fatigue of solder interconnects between each component soldered to a printed circuit board (PCB).  Shock and vibration environments create oscillatory flexure of the printed circuit assembly and create cyclic mechanical stresses in each of the solder joints.  Similarly, thermal cycling between hot and cold temperatures generates stresses as a result of mismatched coefficients of thermal expansion between the PCB and an electrical component.  The PCB and component will expand and contract with temperature at different rates.  Given enough time, these stresses will fatigue the solder interconnects and create an open or intermittent circuit and the system will fail.

 

Analysis Techniques and Problems with the Current Status Quo

Conducting an ERA can include various techniques which usually rely on a mixture of empirical models, Finite Element Analysis (FEA), and Accelerated Lifecycle Testing (ALT).  ALT is very expensive, so the first step is usually empirical model estimates like those developed by Engelmaier, some derived form of Coffin-Manson, or Blattau.  These models:

  1. Force the analyst to make several ambiguous assumptions and input material properties which may or may not be realistic within operating environments.

  2. Generate results which are highly subject to user experience.  Results can drastically over- or underestimate failure prediction because empirical models are often developed and applicable for specific environmental or test conditions and may not generalize well to conditions which fall outside that envelope.

  3. Do not provide any meaningful uncertainty in their estimates.  This is precisely why the space industry analyzes electronic reliability with a safety margin of 400%.  Often times meeting this metric is difficult and results in unnecessary design and analysis iteration and hardware/manufacturing cost if a component just will not pass empirical analysis. 

The next step in the ERA chain is FEA.  FEA can be a great resource, but again, results are highly subject to analyst experience and rely completely on material constitutive relations (how a material behaves in an environment), detailed knowledge of nonlinear material interactions, mesh generation, and model refinement.  Many times the analyst will go through a process of trial and error to find a nonlinear material model which allows the solution to converge.  This inefficient process can generate better reliability estimates than empirical models, but is often expensive and time-consuming.  Also, FEA software is expensive and not all packages are created equal.  An organization may have licenses to several packages depending on how the analysis must be structured.  Learning curves for FEA software can be steep, and customer switching costs to transition to a package more suited to an analysts needs can be substantial.

Finally, ALT is the current "gold standard" in ERA as it represents the hard truth about how well an electronic system will survive an environment and prediction uncertainty is easily determined through statistical methods with a large enough data set.  ALT data is usually expressed as a percentage of failed components in a population versus number of cycles to failure and approximates a Weibull probability distribution, shown in the example graphic below for a single component type and local system.  Engineers and Analysts generally approximate the ALT data with a two- or three-parameter Weibull fit (strait line in graph) to generate their reliability predictions.  But as you can see in the graphic, even this method can drastically over- or underestimate the actual failure data, particularly at the failure extremes.  Additionally, Weibull predictions are not generalized, meaning they are really only useful for a particular PCB layup, component, solder, and other variables that are "baked in" to the data.  Ultimately this means engineers and analysts would have to perform a new ALT series every time a variable in their system changed (e.g. solder or component) because results are highly dependent on even small design variables, like changing from a two millimeter part to a three millimeter part due to availability.

Weibull OverUnder.JPG
 

How PEARL-AI Works and Improves ERA

PEARL-AI is an application built on thousands of ALT data generated by government, private industry, and academia and leverages a proprietary hybrid supervised machine learning model that can make inferences on the complete collection of data to generalize predictions for a wide variety of system features.  It uses over 50 different features as predictors which span the full manufacturing-to-operation lifecycle of the system, ultimately capturing all of the different variables which affect reliability performance from birth until death of the system.  This is a stark contrast to the limited information provided by empirical or FEA estimates and a vast improvement even over predictions generated by Weibull approximations.  Some key facts about PEARL-AI​ are below.

User Friendliness

Using our provided template, thermal cyclic solder fatigue analysis can be conducted in as little as 20 minutes for a complete system.  This saves cost and design time over other methods.

Meaningful Results

Predictions are based on test data.  Period.  These predictions are also accompanied by uncertainty metrics which complete the picture engineers and analysts need to communicate results to customers.

Depth of Data

A typical Weibull fit may include between 20-100 data points.  PEARL-AI leverages over 250000 feature references for predictions. And this training set is growing every day.

Program Specific Tailoring

PEARL does not just spit out numbers.  It does the legwork for you by correlating its results to your program specific environments and design requirements, all in an easy to read format.

The PEARL-AI workflow is shown below.  We provide you with a user-friendly template in Microsoft Excel format.  You fill in all the relevant fields and submit it to the web application portal (see note below).  Then we submit that to the PEARL-AI analysis engine for predictions and it emails you a concise report with all of the information you need.

PEARL-AI Workflow.JPG

Just how accurate is our model?  The below graphic shows the percent difference between PEARL-AI predictions made on a subset of our data and the actual data values.  The x-axis is the quantile of data which falls within the percent difference range.  In other words, PEARL-AI accurately predicted 85 percent of the data subset to within plus or minus 10 percent.  There are always going to be outliers associated with a dataset this size and the fact that this is a highly generalized model which incorporates approximately 30 component categories, 15 solders, 5 PCB finishes, and can even account for thermal/grounding pads and underfills, among the other 45 predictors, makes this a very robust tool.  These error values should also improve as we collect and apply more data, and the data set is growing every day.  Even with all that variability, PEARL-AI still accurately makes 98 percent of its predictions within 42 percent of actual test data.  That is a factor of 10 improvement over the uncertainty adjustment the space industry must make when it analyzes solder interconnect reliability to 400 percent of life.  

PEARL-AI prediction accuracy.png

PEARL-AI is currently built only for thermal cyclic solder fatigue as we are based in Denver, a major space industry hub.  Satellites and flight hardware undergo tens to hundreds of thousands of thermal cycles and we felt this was the most logical place to start based on our relationships with the DoD and space developers.  However, the shock and vibration model architectures have been built and we have amassed a lot of data for use with these models.  We expect the shock and vibration components of PEARL-AI to deploy within the next year or so.  At that point we will update the template, model, correlation, and report and will notify all current licensed users of its availability.  

The web application is also currently under development, but we are capable of making predictions and generating a report for you if you contact us directly.  We expect the web application to be deployed by summer of 2022. In the meantime, we are happy to provide an estimate to do the analysis for you, which we think you will find very reasonable based on our experience of the time and resources necessary to conduct a traditional solder fatigue analysis.