Virtual Glass Testing
Testing Glass Without Breaking Any
A Walkthrough of the Virtual Bend Test — Reader’s Edition
Joseph P. McFadden Sr. • July 07, 2026
Mark10 Glass Flexure Analysis Suite
Hi, my name is Joe McFadden. I’ve spent more than forty years as an engineer, an educator, and an investigator of failures. Over those years I’ve broken a great deal of glass—some of it on purpose, in my home laboratory, on fixtures I printed myself, and some of it the way everyone breaks glass, which is to say by accident, followed by a moment of quiet reflection about what I could have done differently. The slides are numbered—we begin on slide 1, the title.
Slide 2. Here is that setup—a benchtop force frame, fixtures I designed and printed myself, and a lab notebook full of fracture maps. I want you to notice something about those pictures. You do not need expensive equipment to perform complicated testing. You need imagination and foresight—and a willingness to break a few things carefully.
And none of this equipment arrived with a grand plan attached. I originally built that bend fixture—and the software behind it, the Weibull analysis program included—for one simple reason: I wanted to see for myself how glass fractures, and what cutting really does to it—the defects, the wakes of damage the scoring wheel leaves behind. That itch to poke at a question and learn led me down path after path: more fixtures, more software tools, each one built to see a little further into how glass behaves, and how its lived experience makes it what it is at the moment of testing. The Virtual Mark10 you are about to use is simply the latest stop on that road.
In these pages I want to walk you through something a little unusual: a bend test campaign on TFT display glass in which not one specimen gets broken. Not because we’re being careful, but because the specimens don’t exist. We’re going to simulate the entire test—the glass, the machine, the load cell, even the scatter in the results—and I want to spend some time up front on the why behind that, because if you don’t understand why we model, the model is just a toy. If you do understand it, the model becomes one of the best teachers you will ever have.
If you have the slide deck handy, good—the slides are numbered and I’ll call each one out as we go. If you don’t, that’s fine too. The story stands on its own.
Part One — The Nature of Glass
Before we test anything, virtual or otherwise, we need to talk about what glass actually is. Because glass is strange. It is one of the oldest materials we make and one of the least understood by the people who use it every day.
Slide 3. Most solids are crystals. Their atoms sit on a repeating lattice, the same arrangement copied billions of times in every direction, like bricks in a very disciplined wall. Glass is not like that. When molten glass cools, its silica network never gets the time to organize itself into a crystal. The disorder of the liquid simply freezes in place. What you’re left with is an amorphous solid—a material with the rigidity of a solid and the randomized atomic arrangement of a liquid. The TFT glass in your phone and your monitor is an alkali-free aluminoborosilicate, engineered to be electrically clean enough to sit under transistors, but underneath the chemistry it shares this fundamental nature with every glass ever made: it has no lattice.
And that single fact—no lattice—explains almost everything about how glass behaves under load.
Slide 4. Bend a paperclip and it stays bent. The metal deforms plastically, because crystals contain dislocations—line defects that let planes of atoms slip past one another at stresses far below the strength of the bonds themselves. Plastic flow is the safety valve of metals. When the stress at a scratch gets too high, the metal yields locally and blunts the danger.
Glass has no lattice. Therefore no slip planes. Therefore no dislocations. Therefore no safety valve. Glass responds to load in exactly one way: perfectly elastic stretching of its atomic bonds, right up to the instant those bonds break. On a force-displacement plot, glass draws you a straight line from first contact to a vertical cliff. No yield point. No plastic plateau. No warning. Everything about testing glass—and everything in this walkthrough—follows from that one sentence.
Slide 5. Now, before the materials scientists in the audience write me a letter, let me be precise—because precision matters, and this next part is true and worth knowing. On a small enough scale, glass can yield. Press a diamond indenter into a glass surface and you’ll leave a permanent mark: the open silica network compacts—a mechanism called densification—and it shears and flows besides. Every scratch that stays in glass is microplasticity in action. But here’s the catch. That plastic zone is nanometers deep, and with no dislocations to carry the deformation outward, it stays nanometers deep. It cannot blunt a crack. It cannot soak up energy the way a metal does—a crack tip in steel dissipates thousands to millions of times more energy in its plastic zone than glass ever can. So glass yields in nanometers, and parts fail in millimeters. At the scale of anything you would build or hold, glass behaves as perfectly elastic-brittle. That is the sense in which I mean no safety valve.
Now here is where the story gets genuinely strange, and where, as with so many of the failure stories I’ve told over the years, all roads lead back to energy.
Slide 6. Do a simple calculation with me. The strength of a solid ought to be set by the force needed to pull its atomic bonds apart. Work that out and you get roughly E/10. For our TFT glass, with a modulus around 74 GPa, that predicts a strength in the neighborhood of 7,000 MPa. Seven gigapascals. Now take a real strip of that glass, put it in a bend fixture, and load it. It breaks somewhere around 90 MPa. The real material is nearly a hundred times weaker than its own chemistry says it should be.
That paradox sat unresolved until 1920, when Alan Arnold Griffith published what I consider one of the most consequential papers in all of materials science. And I want you to notice what kind of argument Griffith made, because it wasn’t an argument about force. It was an argument about energy.
Griffith said: the culprit is flaws. Microscopic surface cracks, put there by manufacturing, by handling, by dust, by the simple act of existing in the world. Stress concentrates ferociously at a crack tip—the sharper the crack, the higher the local stress—so a plate of glass carrying a modest average stress can be at bond-breaking stress at the tip of one invisible scratch. But here’s the energy part, the part I love. Griffith showed that a crack grows when the elastic strain energy released by its growth exceeds the energy cost of creating the two new surfaces the crack leaves behind. Stored energy in, surface energy out. When that ledger tips in the crack’s favor, the crack runs—and in glass it accelerates toward roughly half the speed of sound. The specimen is in pieces before the load frame registers anything but a cliff in the force channel.
Those of you who’ve read my telling of how fluids kill polymers will recognize this bookkeeping immediately. It is the same energy balance. In the polymer story, a fluid sneaks in and lowers the cost of making new surface, and cracks that were dormant come alive. In glass, the surface energy term is set by the flaw the glass already carries. Different materials, different agents—same ledger. Once you see fracture as energy accounting, you never unsee it.
Slide 7. Griffith proved his point with an experiment I find beautiful. He tested freshly drawn glass fibers, whose surfaces were nearly pristine, and measured strengths approaching that theoretical limit—strengths that then decayed within hours, as the surfaces picked up flaws from the air and from handling. Think about what that means. Strength, in glass, is not a property of the material. It is a property of the material’s worst flaw. The glass carries its history on its surface, the way you and I carry ours—every scratch a lived experience, every lived experience shaping what it can endure next.
Slide 8. And the environment, it turns out, is not a bystander in any of this. Ordinary humidity—the water in the air of every lab and every living room—finds its way to the tip of a crack. And at that tip, where the silicon-oxygen bonds are stretched taut by the load, a water molecule can do something it cannot do anywhere else on the glass: it reacts with the strained bond and cuts it. Chemists call it hydrolysis. I call it Griffith’s ledger with a corrosive accountant on the payroll. The crack advances one broken bond at a time, slowly, patiently, at loads far below what fast fracture would demand. We call it subcritical crack growth, or stress corrosion. And if you have read my telling of how fluids kill polymers, you have met this villain before—an environment quietly lowering the price of new surface until a crack that had no business moving begins to move.
Slide 9. The consequences show up at the test frame in two ways, and both matter to everything we do later. First, static fatigue: a piece of glass that survives a load today can fail next month under that very same load, because the crack has been creeping the whole time. Survival is not a property; it is a race between the crack and the calendar. Second, loading rate: test the same glass slowly, and water has time to work, so it measures weaker. Test it fast, and you outrun the corrosion, so it measures stronger. The same material gives you different answers depending on how quickly you ask the question. That is why the standards specify a loading rate, why the Virtual Mark10 asks you to choose one, and why a strength quoted without its rate and its environment is an incomplete sentence. One honest note about our simulator: the virtual glass does not yet model stress corrosion—it fails the same way at any rate. That is a simplification, and knowing your model’s simplifications is part of using models well.
Slide 10. And if strength is set by the worst flaw, then strength cannot be a number. Two specimens cut from the same sheet carry different flaw populations, so they break at different loads. A batch of glass specimens is a batch of chains, each failing at its weakest link, and the weakest link differs from chain to chain. Ask “how strong is this glass?” and the honest answer is a distribution, not a value.
Slide 11. Waloddi Weibull gave that distribution its standard mathematical form, with two parameters you’ll estimate yourself before this walkthrough is over. The characteristic strength η locates the distribution: it’s the stress by which 63.2% of specimens have failed. The Weibull modulus m measures the scatter. High m means a uniform flaw population and predictable failures. Low m means a wild flaw population and enormous spread. Ordinary glass resides down around m = 6—among the most scattered materials in all of engineering. Our TFT glass, made under cleaner conditions, does a bit better, around m = 8. Still scattered. Still a distribution. Still glass.
Slide 12. Let me make this concrete with a story from my own bench—the same bench in those photographs. Thin glass gets cut by scoring and snapping, and the scoring wheel is not gentle. Under the microscope, a scored edge shows a procession of median cracks—little arcs marching along the edge where the wheel rolled, each one a flaw that the cutting process wrote into the glass. Now, if strength is the property of the worst flaw, then it should matter enormously where those cracks end up when you load the part. So I sorted my specimens into two groups: one with the scored edge on the tension side of the bend, and one with it on the compression side. In three-point bending, the bottom surface carries the tension—put the median cracks down there, and you are loading your worst flaws directly. My early results say exactly what the theory predicts: score in tension, and the glass fails at markedly lower loads. I’ll present that data properly in another telling, because right now I do not have enough specimens for honest statistics—and after everything I’ve just told you about scatter, I hope you understand why I refuse to draw conclusions from a handful. Some time in the future I’ll have enough for proper sampling. Patience is part of the method.
KEY LESSON: One measurement of glass strength tells you almost nothing. Hold onto that sentence. It is the reason everything that follows involves twenty specimens instead of one.
Part Two — The Why Behind the Model
Slide 13. So why simulate? I built a virtual test machine into my analysis software—a Virtual Mark10, modeled on the real force frame that sits in my home lab—and I want to be honest with you about why, because “it’s convenient” is not the reason.
Here is the problem with learning from real test data: nobody knows the right answer. You run twenty specimens, you compute a modulus, you fit a Weibull distribution—and then what? Is 72 GPa correct? Is your Weibull modulus of 7 the truth, or an artifact of your sample size? In a real lab, there is no answer key. You can be confidently, precisely wrong and never find out. I have watched that happen in industry more times than I care to count, and the failures that follow are rarely gentle.
A simulation flips that on its head. In the virtual machine, I know the truth going in, because I put it there. The material in the library has an exact modulus, an exact characteristic strength, an exact Weibull m. Every specimen the simulator creates obeys real beam mechanics—the same equations your professor wrote on the board—with realistic machining tolerance on its dimensions and a failure stress drawn honestly from the material’s Weibull distribution. The simulated load cell quantizes force the way the real sensor does, saturates at the same 50 N, and keeps recording zeros after the specimen breaks, just like the real frame. The files it writes are byte-for-byte indistinguishable from real machine output. Which, by the way, is exactly why I’ll remind you later to label your folders. A simulation good enough to teach from is a simulation good enough to fool you.
Because the truth is known, every step of your analysis can be graded. Predict, generate, analyze, compare. That closed loop is something a real laboratory can never quite offer. Master the loop in simulation, and real data will hold few surprises. That’s the why. It’s the same holistic instinct that runs through everything I teach: understand the whole chain—the material, the machine, the data, the statistics—because failures reside in the seams between those things, in the places where one specialist’s knowledge ends and another’s hasn’t started.
And there’s one more reason, a personal one. The most valuable habit I know in experimental work is predicting the answer before you take the data. Not guessing—predicting, with equations, on paper. When your prediction and your measurement agree, you’ve earned trust in both. When they disagree, you’ve found something worth your attention—a mistake, a unit error, an artifact, or occasionally, if you’re very lucky, something new. The virtual machine lets a student practice that habit fifty times in an afternoon without consuming a single strip of glass.
Part Three — The Walkthrough
All right. Let’s run the campaign. Our specimen is a strip 0.25 mm thick—250 micrometers, thinner than three sheets of paper—38 mm wide, resting on supports 20 mm apart. Real display-glass dimensions, the kind I handle with the scars to prove it.
Slide 14 — The plan, and the truth
Slide fourteen lays out the plan, and it is exactly the loop I described: predict, generate, analyze, compare. On the right of that slide you’ll see the truth we’re testing against, straight from the material library: modulus 73.6 GPa, characteristic strength η = 90 MPa, Weibull m = 8. Write those three numbers down. At the end of this walkthrough, you will grade yourself against them.
Slide 15 — Predict before you press any button
Three lines of beam theory, and I want to walk through them slowly because these three lines are the whole mechanical story.
First, the second moment of area. Notice the thickness is cubed—tuck that away; it comes back to haunt us in the best possible way.
I = b·h³/12 = 38 × 0.25³ / 12 ≈ 0.0495 mm⁴
Second, the stiffness. For three-point bending, the slope of force against deflection is set by the modulus and the geometry—and that is the slope every single specimen should draw for us, because stiffness comes from the atomic bonds and the geometry, and those barely vary.
k = 48·E·I/L³ = 48 × 73,600 × 0.0495 / 20³ ≈ 21.9 N/mm
Third, the failure. Take a typical strength around 85 MPa—the mean of a Weibull distribution with η = 90 and m = 8:
F = 2·σ·b·h²/(3L) ≈ 6.7 N, δ = F/k ≈ 0.31 mm
A third of a millimeter. That’s the entire test. So our prediction: straight lines near 22 N/mm, breaking somewhere around 5–9 N, all comfortably inside our 50 N load cell. The span matters here, by the way—L is cubed too. At a 30 mm span these same strips would be nearly 3.5× softer. Geometry is never a detail in bending. Geometry is most of the story.
Slides 16 and 17 — Setting up the machine
In the software, it’s the Tools menu, then Generate Test Data—the Virtual Mark10. The dialog is laid out the way you’d set up the real frame, top to bottom. Material first: TFT display glass, alkali-free, and notice the dialog shows you the library truth right beside the selector—the same three numbers you wrote down. Then geometry: span 20, fixed for the whole batch. Width 38 ± 0.10 mm. Thickness 0.25 ± 0.005 mm. Those tolerances aren’t decoration—the simulator treats them as a 3-sigma machining band and gives every specimen its own slightly different dimensions, exactly like a real batch coming off a cutter.
Slide 18 — The machine settings, and the why behind each
Pusher travel, 5 mm—our breaks are predicted near a third of a millimeter, so 5 mm guarantees every specimen fails, and the machine records zeros afterward just like the real IntelliMESUR does. Loading rate, 30 mm/min. And here’s a piece of test planning I want you to catch: at this short, stiff span, the whole test is over in a third of a millimeter, so if we ran at the usual 60 mm/min with 100 Hz sampling, we’d collect barely thirty data points before fracture. Thin data makes shaky fits. So we halve the rate, get a point every 5 µm of stroke, and collect about sixty points to failure. Sizing the test rate to the specimen—that’s the kind of decision that separates a measurement from a number. Finally, twenty specimens, a fixed random seed so the dataset is reproducible, load-cell noise on because reality is noisy, and the instructor answer key on—though if you’re a student, that file is not for you yet.
Slide 19 — Generate and load
Click generate, pick a folder, and twenty files appear in native Mark10 format, plus the answer key. The software offers to load them straight into the analyzer—say yes—and then enter the same nominal geometry in the main window. The same geometry. The simulator varied each specimen slightly, but you, like a real technician, measure once and enter one set of numbers. The scatter that causes is honest, and we’ll meet it again in a moment. And please—label that folder as simulation. These files are indistinguishable from real machine output. Future you will thank present you.
Slide 20 — Twenty curves, one story
Process the data, open the force-displacement plot, and just look at it for a moment. Twenty curves. Every one of them rides the same slope—right on the beam-theory line at 21.85 N/mm. And every one of them ends at a different height, from about 5.2 N up to 8.7 N. One stiffness. Twenty different endings. If Part One did its job, you’re not surprised—you’re nodding. Stiffness resides in the atomic bonds, which every specimen shares. Strength resides in the worst surface flaw, which every specimen carries alone. That single picture is the nature of glass, drawn by the data itself.
Slide 21 — From slope to modulus
The software picks a fit region automatically—skipping the very first points and the break—and regresses force against displacement through the clean elastic middle. For specimen one: slope 21.66 N/mm, which the beam equations turn into a modulus of 73.0 GPa, with R² = 0.997. And a caution from an old investigator: a beautiful R² tells you your fit region is internally consistent. It does not tell you the region is right. Glass will happily give you a gorgeous straight-line fit through a fixture artifact. Judging where the physics starts is your job, not the regression’s.
Slide 22 — Grading time
Across all twenty specimens, the measured modulus came back at 73.29 GPa against a truth of 73.6. Off by four tenths of one percent. The coefficient of variation was 1.4%—and here’s the question I promised: where does that scatter come from, when every simulated specimen has exactly the same modulus? It comes from geometry. Remember the thickness, cubed, in that first equation? A 2% thickness tolerance becomes roughly 6% in stiffness, specimen to specimen, and you divided every measurement by one nominal geometry. The scatter in your modulus column is machining, not material. In forty years of failure work, I cannot count the times someone chased a materials problem that was actually a dimensions problem. The mean strength came in at 89 MPa—right in line with what a Weibull distribution of η = 90, m = 8 should hand a sample of twenty. And the worst R² in the whole batch was 0.995. Glass is nothing if not linear.
Slide 23 — The Weibull view
The most important slide in the deck, I’d argue. The histogram shows what strength really is—a spread from the low sixties to over a hundred megapascals, a 14% coefficient of variation—and the Weibull plot straightens that spread into a line whose slope is m. Our fit came back with m = 7.8 against a truth of 8—within 3%, a genuinely lucky draw—and η = 94 against 90, 5% high. And that is with clean, well-behaved simulated data. Twenty specimens estimate a scatter parameter only roughly; run the same campaign with a different seed and watch m wander. This is precisely why the testing standards ask for thirty or more specimens when the Weibull parameters are the deliverable, and why one broken strip of glass tells you almost nothing at all.
Slide 24 — Where to take it next
Four suggestions, in rising order of mischief. Turn on the fixture-settling artifact and learn to spot the compliant toe it puts at the start of every curve—then learn to exclude it, which is the single most valuable habit in flexure testing. Change the span to 30 mm and verify the stiffness drops to about 30%, because L is cubed and the equations mean what they say. Grow the sample—ten, twenty, fifty specimens—and watch the fitted m converge on 8, feeling in your hands why the standards demand the numbers they demand. And finally, play glass detective: have a partner generate a dataset from a library glass they won’t name, and identify it from stiffness and scatter alone. Stiffness tells you the family. Scatter tells you the surface. Together, they tell you the glass.
Closing Thoughts
Let me leave you with the thought underneath all of this. We simulated a bend test today, but what we really practiced was a way of thinking: predict from first principles, measure honestly, compare against truth, and treat every disagreement as a teacher. The glass, for its part, told us its whole story in two numbers—a stiffness set by bonds it was born with, and a strength set by the flaws it picked up along the way. Its lived experience, written on its surface, read back to us through a load cell.
And once again, quietly, underneath the mechanics, it was an energy story—stored elastic energy on one side of Griffith’s ledger, the cost of new surfaces on the other, and fracture the moment the books tip. The same accounting that governs a cracked polymer pipe, a skater on thin ice, and a strip of display glass a quarter millimeter thin.
Now go break some virtual glass. It’s the only kind that doesn’t leave scars. Thanks for reading—I’m Joe McFadden.
