The challenge stems from the fundamental architecture of FEA. Traditional solvers require the stent’s geometry to be subdivided into elements, closed one-, two-, or three-dimensional shapes composed of nodes connected by lines or curves. Each element contains one or more integration points where strain and stress are actually calculated. When a stent frame bends or twists, the peak strains and stresses occur at the structure’s surface, yet the integration points of three-dimensional elements sit inside the element volume, away from that surface. Refining the mesh, meaning a uniform reduction in element size, pushes those integration points closer to the surface and generally improves accuracy. The catch is that finer meshes demand dramatically more computational resources, and engineers must justify every hour of simulation time against the marginal gains in numerical precision.

The industry’s traditional answer has been a mesh refinement study reporting the fractional change in a quantity of interest, typically peak strain, between successive refinements. A common acceptance criterion for medical device simulations is a fractional change of 5.0 percent or less between peak strain predictions. But the authors argue this threshold is arbitrary and disconnected from the real stakes. As they point out, the 5.0 percent criterion bears no relation to the patient or business risk associated with making a decision based on the simulation results, nor to the computational cost of the simulations. Worse, the fractional change calculation provides no insight into the actual error and uncertainty in the prediction due to the mesh’s discretization, leaving analysts without the evidence needed to defend their model choices before regulators or internal review boards.

To close this gap, the team, drawing on engineers from Gore Medical, Medtronic, Ansys, Becton Dickinson, and Nuno Rebelo Associates, built a methodology anchored in three pillars: discretization error, computational cost, and model risk. The approach begins by estimating the exact solution of the numerical model, the value the simulation would converge to if the mesh could be refined indefinitely. They employed Richardson extrapolation, which estimates the converged value from a sequence of systematically refined meshes, and compared two ways of quantifying uncertainty. The first is the Grid Convergence Index, or GCI, a well-established technique that calculates numerical uncertainty with a safety factor, typically 1.25 or 3.0 depending on how closely the observed order of accuracy matches the theoretical value. The second is a newer method proposed by P. Krysl in 2022, which requires four systematically refined meshes rather than three but dispenses with the variable safety factor entirely.

The Krysl method proved the authors’ preference. It works by calculating the convergence rate for all four possible combinations of three meshes drawn from the four refinement levels, extrapolating a solution for each combination, and taking the median of those extrapolated values. The estimated exact solution is then a weighted combination, in a 2:1 ratio, of the extrapolation from the three finest meshes and the median of all four. Uncertainty is quantified using the median absolute deviation, scaled by a constant of 1.4826 under the assumption of a normal distribution. In the team’s three-point bend benchmark, the GCI method using the three finest meshes estimated an exact solution of 0.008006 with a 95 percent confidence interval spanning 0.000616, while Krysl’s method, using all four meshes, estimated 0.008000 with a far narrower interval of just 0.000080. Both methods placed the confidence band around the estimated exact solution itself, but Krysl’s tighter range gave analysts greater credibility per unit of computational effort.

To demonstrate real-world applicability, the team applied the methodology to a laser-cut nitinol stent frame, the superelastic nickel-titanium alloy that dominates modern stent design. The test article was a strut section cut from a nitinol tube with a 1.2 millimeter outer diameter and 0.1 millimeter thickness. Using symmetry, the model was reduced to a single strut, which was then simulated through a complete lifecycle: expanded to a 4 millimeter outer diameter from its as-cut geometry, annealed to mimic the shape-setting step in manufacturing, radially compressed to 1.5 millimeters to simulate catheter loading, and finally released to 3.2 millimeters to represent deployment into an artery 20 percent smaller than the stent’s nominal size. Submodeling of the region of interest kept the mesh refinement study affordable, with meshes containing 4, 8, 16, and 32 elements across the strut width systematically refined by a factor of two at each level.

Crucially, the team ran the entire exercise across four different finite element solvers, anonymized as solvers A through D, spanning products from Dassault Systèmes and Ansys. The converged solutions differed between solvers, an expected outcome given variations in physics, constitutive equations, element formulations, and boundary constraint methods, but the methodology itself proved agnostic to solver type and developer. For each solver, Krysl’s method produced an estimated exact solution with uncertainty, such as 0.010333 plus or minus 0.000136 for solver A. The team also tested whether expensive quadratic elements were necessary, running a three-point bend benchmark with four hexahedral element types. All four converged toward a peak maximum principal strain of approximately 0.008, showing that a linear element with a single integration point can be as accurate as a quadratic element with over 20 integration points, though the linear reduced-integration element required roughly four times as many elements through the beam thickness to match the quadratic element’s prediction.

With the exact solution estimated, discretization error and computational cost could be calculated for candidate meshes. The team evaluated meshes with 6 and 8 elements across the strut width at a 1.0 aspect ratio, and 8 and 10 elements at a 1.25 aspect ratio, where the element lengths in non-critical directions were set 25 percent greater than in the critical width direction to trim element counts without sacrificing density where the strain gradient is highest. The results revealed genuine trade-offs. The six-element mesh carried discretization errors between 4.2 and 9.8 percent across solvers but ran at 0.4 to 0.8 times the computational cost of the reference eight-element mesh. The ten-element, 1.25 aspect ratio mesh reduced error for three of four solvers, down to as little as 0.3 percent, but at a cost of 1.0 to 1.3 times the reference. No single mesh emerged as universally correct.

That subjectivity is precisely the point, and the authors grounded it in the ASME V&V 40 standard, which assesses model credibility by considering the influence of simulation results on a decision and the ramifications of an incorrect one. They presented three realistic contexts of use spanning a continuum of model risk. In the low-risk scenario, screening 50 stent designs across three boundary condition sets with no patient risk, the team selected the computationally efficient eight-element, 1.25 aspect ratio mesh with a 5.0 percent discretization error, prioritizing cost across roughly 150 simulations. In the medium-risk scenario, certifying fatigue requirements for a stent family where the simulation informs subsequent in vitro testing, they chose the ten-element mesh with 4.1 percent error, accepting a 1.3 times cost because the model’s output guides physical testing. In the high-risk scenario, where the computational model is the sole means of assessing durability with no in vitro fatigue testing planned, the same ten-element mesh was selected despite each of twelve simulations expected to exceed 40 hours, because the elevated stakes demanded the lowest achievable discretization error.

The authors deliberately declined to set a universal discretization error acceptance threshold, reasoning that the contribution of numerical uncertainty to a model’s overall accuracy cannot be determined during the mesh refinement study alone, since total uncertainty also includes input uncertainty and model form error. Instead, they recommend selecting the mesh based on the calculated discretization error, the computational cost, and knowledge of the risk associated with using the model to inform the decision, with a poor match against validation data signaling the need for further refinement. For an industry in which FEA evidence shapes regulatory submissions and patient outcomes alike, the methodology offers something the old 5 percent rule never could: a defensible, transparent, and risk-aware rationale for every mesh choice, one that two analysts can debate openly rather than guess at silently.

Subject of Research: Mesh discretization selection methodology for finite element analysis of cardiovascular stent frames

Article Title: Finite element analysis of cardiovascular stent frames: identifying appropriate mesh discretization

Article References: Lurie, B. A., Kari, S., Horner, M., Basciano, C. A., Kumar Murugadass, S., & Rebelo, N. (2025). Finite element analysis of cardiovascular stent frames: identifying appropriate mesh discretization. Frontiers in Medical Engineering, 3, Article 1606951. https://doi.org/10.3389/fmede.2025.1606951

Image Credits: AI Generated

DOI: 10.3389/fmede.2025.1606951

Keywords: finite element analysis, cardiovascular stent, mesh discretization, discretization error, model risk, nitinol, Richardson extrapolation, Grid Convergence Index, Krysl method, ASME V&V 40, submodeling, medical device simulation

The research, carried out by Samuel Onimpa Alfred of the Department of Aerospace Engineering at the University of Michigan, belongs to a rapidly growing field called triboinformatics, where artificial intelligence is used to decode the messy, nonlinear world of friction, lubrication, and wear. Tribology, the science of interacting surfaces in relative motion, is not a niche pursuit. Earlier studies cited in the work estimate that friction and wear-related losses consume a staggering share of global energy, contributing substantially to costs and emissions across industry. Making surfaces last longer, in artificial hip joints or in jet engine bearings, translates directly into energy saved and devices that survive longer inside the human body or inside an aircraft engine.

What makes Nitinol such a fascinating test case is its shape-memory and superelastic character. Deform it and it snaps back; heat it past a critical temperature and it returns to a previously memorized configuration. But these same properties make its tribological behavior notoriously complex. In this study, Alfred examined four distinct Nitinol conditions: an equiatomic titanium-nickel alloy, and three versions of a nickel-rich composition known as 60NiTi that had been aged, annealed, or solution-treated. Each heat treatment changes the alloy’s microstructure, hardness, and elasticity, and therefore changes how it wears when scraped against a counterface in dry, reciprocating sliding. Classical wear laws, which tend to assume wear scales simply with load and sliding distance while inversely scaling with hardness, capture these shifts poorly.

To tame that complexity, Alfred assembled a dataset of 336 individual measurements drawn from previously published, peer-reviewed reciprocating dry-sliding experiments, covering weight loss, cumulative wear over time, and steady-state coefficient of friction for all four material conditions. The modeling strategy then subjected a battery of algorithms to a demanding examination. Instead of splitting the data randomly, the study used grouped cross-validation, meaning that entire series of wear tests were withheld from the model during training. This is the scientific equivalent of asking a student to answer questions about chapters of a book they were never allowed to read, and it guards against the overly optimistic predictions that plague many machine-learning studies in materials science.

Even under this harsh test, gradient boosting, an ensemble method that builds a predictive model from many sequentially corrected decision trees, emerged as the clear winner. It predicted weight loss with an R-squared of 0.92 and the coefficient of friction with an R-squared of 0.91, beating support-vector regression, random forests, and simple linear baselines. When the cross-validation was made less restrictive and random splits were allowed, those scores climbed to 0.98 for wear and 0.94 for friction. The result confirms that gradient boosting does not just memorize data; it learns genuinely transferable relationships between operating conditions and tribological outcomes, even for material-test combinations it has never encountered.

Crucially, the study did not stop at prediction. Using SHapley Additive exPlanations, or SHAP, a technique borrowed from cooperative game theory that assigns each input variable its fair share of credit for a model’s output, Alfred opened the black box. The analysis revealed that applied load is the dominant driver of wear, while oscillation frequency dominates friction behavior, with higher frequencies associated with lower friction coefficients, a trend consistent with frictional heating at the sliding interface. This kind of transparency matters enormously for engineers, because a model that merely outputs numbers without explanations offers no guidance on which design levers to pull.

Perhaps the most striking achievement is a deep learning result: a gated recurrent unit network, a type of neural network designed for sequential data, reproduced the full time-resolved wear trajectories of completely unseen tests with an R-squared of 0.91. In other words, given the early portion of a wear test, the network could accurately trace how material loss would accumulate over the entire remaining test, test after test, across all four heat treatments. That capability opens the door to digital wear forecasting, where a short initial experiment or monitoring window could stand in for long and expensive laboratory campaigns.

The study then confronted the granddaddy of wear science, the Archard equation, formulated in 1953, which states that wear volume is proportional to load and sliding distance and inversely proportional to hardness. When that classical law was fitted to the Nitinol dataset, it managed an R-squared of only 0.27, and performed even worse, going negative, when hardness was imposed rather than fitted. Alfred instead let the data speak, recovering a generalized Archard-type law in which the exponents on load, sliding distance, frequency, and hardness are free parameters. The resulting equation, weight loss equals 0.0425 times load to the power 0.78, sliding distance to the power 0.54, frequency to the power minus 0.14, and hardness to the power minus 0.26, described all four heat treatments with an R-squared of 0.79, using only measured hardness rather than material identity labels.

The exponents themselves tell a physical story. The sub-linear load exponent of 0.78 suggests that superelastic Nitinol distributes contact stress in a way that softens the wear increase as loads climb, while the negative frequency exponent quantifies the frictional-heating effect seen in the SHAP analysis. Most intriguingly, the fitted wear coefficient for each material correlated almost perfectly, at r equals 0.97, with the ratio of elastic modulus to hardness, a dimensionless quantity long championed in surface engineering as an indicator of elastic, wear-tolerant contact. When the modulus-to-hardness ratio served as the sole material descriptor in the generalized law, the fit reached an R-squared of 0.83, essentially matching models that knew which alloy they were dealing with.

For a metal that must survive inside arteries, bearings, and aerospace mechanisms without the benefit of lubrication, these findings provide something genuinely new: accurate, transparent, and physically consistent models that connect processing, properties, and performance. A designer can now estimate how a given heat treatment, hardness, and duty cycle will translate into wear and friction, before a single prototype is machined. More broadly, the work is a template for how triboinformatics should be done, with strict grouped validation, interpretable explanations, and laws recovered from data that honor the physics of contact rather than discarding it. As Nitinol finds its way into ever more demanding applications, the machines that predict its wear are, fittingly, learning from the metal that never forgets.

Beneath the headline results lies a dataset with an unusually clean provenance. The 336 measurements were not generated afresh for the modeling study but were compiled from two previously published, peer-reviewed experimental campaigns on superelastic TiNi and 60NiTi, with every table—loads, frequencies, durations, sliding distances, specimen masses before and after testing, cumulative weight loss, and steady-state friction coefficients—reproduced in the new paper’s appendix. That decision to expose the full experimental record alongside the models is itself a small contribution to a field where data scarcity and fragmentation remain the chief obstacles to progress.

The geometry underlying those tables also rewards a closer look. Each test used a 5.03 millimeter reciprocating stroke, so every cycle covered just over a centimeter of sliding, and the sliding distances reported for all four material conditions satisfy an exact arithmetic relation linking distance to frequency and test duration. All fifty-six wear-time series were strictly monotonic, with no missing entries, meaning the recurrent network tasked with reconstructing wear trajectories never had to impute gaps—an often unappreciated advantage when deep learning meets sparse laboratory data.

The strong correlation between the fitted wear coefficient and the elastic-modulus-to-hardness ratio also has a pedigree worth noting. Surface engineers have argued for decades that this ratio, rather than hardness alone, governs how well a material tolerates elastic contact and resists abrasion, particularly for coatings and for alloys whose elastic resilience absorbs deformation that would otherwise be permanent. The Nitinol results give that long-standing heuristic a quantitative, data-driven endorsement for shape-memory metals specifically.

The study situates itself in a broader movement. Recent systematic reviews of machine learning in tribology have catalogued a wave of applications, from aluminum-matrix composites to modified zinc alloys to diamond-like carbon coatings, where algorithms predict friction and wear from operating parameters. What distinguishes the present work within that wave is its insistence on withholding whole test series during validation and on recovering an interpretable wear law from the same data used to train opaque models. The author notes that the code underlying the analysis is available on reasonable request, and the supplementary data file, roughly 800 kilobytes in spreadsheet form, invites others to replicate or extend the models for their own shape-memory alloy systems.

Subject of Research: Machine learning prediction of wear and friction behavior in heat-treated Nitinol shape-memory alloys.

Article Title: Triboinformatic modeling of nitinol alloys under different heat-treatment regimes

Article References: Alfred, S. O. (2026). Triboinformatic modeling of nitinol alloys under different heat-treatment regimes. Journal of Materials Science. https://doi.org/10.1007/s10853-026-13713-9

Image Credits: AI Generated

DOI: 10.1007/s10853-026-13713-9

Keywords: Nitinol, tribology, machine learning, wear prediction, shape-memory alloys, gradient boosting, SHAP, Archard equation, deep learning, heat treatment, friction, triboinformatics

The technology’s lineage traces back to the first-generation Merci retriever, a corkscrew-like device approved in the mid-2000s that proved mechanical clot removal in the brain was feasible but achieved only modest recanalization rates. The decisive turning point came with self-expanding stent retrievers, notably Solitaire and Trevo, which are compressed inside a microcatheter, navigated through the vasculature to the clot, and then unsheathed so the nitinol mesh expands and integrates with the thrombus. Unlike a static stent left in place, these retrievers engage the clot mechanically across its full length, allowing the operator to pull the entire construct into a guide catheter. Randomized trials published in 2015, including MR CLEAN, ESCAPE, EXTEND-IA, SWIFT PRIME, and REVASCAT, together demonstrated overwhelming benefit of endovascular thrombectomy over medical therapy alone, reshaping international guidelines almost overnight and establishing first-generation stent retrievers as the anchor of modern stroke intervention.

The technical principle behind the second-generation devices is deceptively simple: radial force from the expanding mesh compresses the clot against the vessel wall while individual struts penetrate the thrombus, creating a mechanical interlock. In practice, the interaction is governed by a complex interplay of clot composition, device geometry, and vessel size. Ischemic stroke thrombi vary enormously, from soft, red-cell-rich emboli shed from the heart to hard, fibrin-rich clots loaded with platelets and von Willebrand factor that resist mechanical integration. Histological analyses cited in the review show that fibrin-dense outer shells and platelet-rich regions correlate with failed retrieval and poorer revascularization outcomes. This biological heterogeneity has pushed engineers toward devices with segmented designs, larger mesh cells, and specialized capture zones. Multi-zone platforms such as NeVa incorporate discrete drop zones with tightly spaced struts designed to trap organized clots, while radially adjustable retrievers such as Tigertriever allow the operator to expand the device progressively until it matches the vessel diameter, an advantage in both oversized proximal vessels and narrow distal branches.

Device development has also converged on integrated retrieval-and-protection concepts. EmboTrap-class retrievers feature distal capture baskets intended to intercept fragments that would otherwise migrate downstream and cause new infarcts, a complication known as distal embolization. registries such as ARISE II and the global EXCELLENT registry for the EMBOTRAP device have reported high first-pass reperfusion rates with these hybrid designs. First-pass effect, meaning complete reperfusion achieved in a single retrieval attempt, has emerged as a key performance metric because each additional pass increases procedural time, trauma to the endothelium, and the risk of hemorrhagic transformation. Recent generation devices such as Solitaire X have demonstrated significantly improved first-pass success compared with their predecessors, and tip-design studies confirm that the shape and stiffness of the retriever’s distal end materially influence whether fragments escape during withdrawal.

In parallel with hardware evolution, procedural technique has advanced into highly choreographed combinations. The most influential refinement is the pairing of stent retrievers with large-bore aspiration catheters positioned at the face of the clot, a strategy variously branded as Solumbra, SAVE, or ARTS. Aspiration continuously extracts clot fragments dislodged by the retriever, reducing the shower of emboli that would otherwise travel into healthy territory. Balloon guide catheters add a second layer of protection by temporarily arresting antegrade flow in the parent artery, creating a stagnant zone from which debris can be vacuumed rather than washed distally. The randomized PROTECT-MT trial from China showed that balloon guide catheters significantly improve excellent reperfusion rates, validating what in-vitro flow studies had long predicted. For refractory occlusions, operators increasingly deploy double stent retrievers simultaneously, doubling the mechanical interface with the clot, and recent bench studies plus the randomized TWIN2WIN trial support this bail-out strategy, although cumulative vessel wall injury remains a documented concern in animal models.

The clinical indications for thrombectomy have expanded dramatically alongside the devices themselves. Landmark trials including DAWN and DEFUSE 3 extended the treatment window from six hours to twenty-four hours in patients selected by advanced perfusion imaging, demonstrating that brain tissue can remain salvageable long after symptom onset when collateral circulation is robust. More recently, attention has turned to posterior circulation strokes caused by basilar artery occlusion, which are uniformly devastating without treatment; trials such as ATTENTION and BAOCHE provided the first randomized evidence supporting endovascular therapy in this territory. Equally consequential are the new studies in large infarct cores, including SELECT2, ANGEL-ASPECT, and RESCUE-Japan LIMIT, which overturned the long-held exclusion of patients with extensive established damage and showed net functional benefit from thrombectomy even in these high-risk presentations.

The most recent frontier involves medium and distal vessel occlusions, blocks in arteries one to three millimeters in diameter that were historically managed with medication because standard devices were too bulky. Purpose-built low-profile retrievers, including 3-millimeter variants of Solitaire X and Trevo and the adjustable Tigertriever 13, have enabled operators to reach these small vessels, and a wave of randomized trials in 2025 and 2026, including DISTALS, DISTAL, and DISCOUNT, has begun to establish benefit under imaging-guided selection. The review emphasizes that territory-specific engineering, from smaller delivery profiles to softer, more flexible distal architectures, is now the dominant axis of device innovation, with hybrid devices such as Aperio and specialized platforms for cerebral venous sinus thrombosis broadening the field further.

Materials science is contributing a quieter but potentially transformative layer of progress. Nitinol remains the workhorse alloy because its superelasticity allows dense crimping and atraumatic self-expansion, but its poor radiographic visibility complicates positioning, prompting coatings and design changes that enhance fluoroscopic contrast. Surface engineering aims to reduce thrombogenicity and endothelial damage, with heparin-based hydrogel coatings, endothelium-mimicking bioactive layers, and nanostructured oxide films under investigation. More provocative are clot-adhesive coatings that deliberately bind to fibrin, effectively welding the retriever to resistant thrombi, and micro-patterned surfaces that increase contact area. In a striking departure from conventional designs, milli-spinner thrombectomy, reported in Nature in 2025, uses a rotating, tangle-forming structure to compress and extract clots regardless of composition, hinting that the retrieval paradigm itself may not be permanent.

Looking forward, the review identifies thrombus characterization as the bridge between biology and device choice. Radiomic analysis of clot appearance on imaging, combined with biomarkers of clot composition, could soon allow operators to predict before the first pass whether a given occlusion will yield to a standard retriever or demand an adjustable device, dual-stent technique, or direct aspiration. Personalized device selection of this kind would attack the core unresolved problems: fibrin-rich resistant thrombi, embolic complications, vascular injury from repeated passes, and the limited high-level evidence supporting many of the newest devices, which have largely been validated in registries rather than randomized trials. The authors argue that ongoing integration of engineering innovation with rigorous clinical data will support increasingly individualized and safer thrombectomy strategies, and with stroke remaining a leading cause of disability globally, even incremental gains in first-pass success translate into meaningful reductions in death and dependence. The stent retriever, born from a simple wire mesh, continues to evolve into a precision instrument tailored to the specific clot, vessel, and patient standing between a stroke and recovery.

The stakes of these technical refinements are best understood against the sheer scale of the disease. Global burden analyses cited in the review estimate that stroke affected well over a hundred million people worldwide in recent years, and large vessel occlusions contribute disproportionately to death and dependence because the entire territory of a major cerebral artery is threatened within minutes of onset. Intravenous thrombolysis, the other pillar of acute reperfusion therapy, dissolves clot biochemically but achieves recanalization in only a minority of large vessel occlusions and carries a risk of arterial reocclusion, which is why mechanical retrieval became indispensable.

The review also situates current practice within the 2026 American Heart Association and American Stroke Association guideline for early management of acute ischemic stroke, reflecting how trial evidence is rapidly codified into standards of care. Beyond the procedure itself, the authors note that reperfusion initiates a second wave of injury, including blood-brain barrier breakdown and neuroinflammation, meaning that restoring flow is necessary but not always sufficient for good functional recovery. This biological reality underscores why procedural metrics such as first-pass success and reduced embolization matter clinically, and why the field increasingly views mechanical thrombectomy not as an isolated engineering problem but as one component of a broader effort spanning imaging selection, device design, and post-reperfusion neuroprotection.

Subject of Research: Technological advances and clinical applications of stent retrievers in endovascular thrombectomy for acute ischemic stroke.

Article Title: Stent retrievers for acute ischemic stroke: technological advances, clinical applications, and future perspectives

Article References: Xie, Z., Wang, Z., Fu, P., Shi, Z., Zhuang, Z., Wang, H., Xiang, Y., Yin, X., & Chen, Z. (2026). Stent retrievers for acute ischemic stroke: technological advances, clinical applications, and future perspectives. Journal of Neurology, 273(10), Article 590. https://doi.org/10.1007/s00415-026-14126-z

Image Credits: AI Generated

DOI: 10.1007/s00415-026-14126-z

Keywords: acute ischemic stroke, stent retriever, mechanical thrombectomy, large vessel occlusion, endovascular treatment, first-pass reperfusion, nitinol, distal embolization, balloon guide catheter, medium vessel occlusion, thrombus composition, reperfusion