15.4 Interpreting Real Motion in Laboratory and Engineering Contexts
The Accelerometer Paradox
[Video: A smartphone is held at chest height over a bridge railing. The screen displays a real-time accelerometer app, showing a steady reading near $9.8\ \text{m/s}^2$ in the vertical axis. A hand releases the phone. The camera (a second phone, mounted on the railing) tracks the phone as it falls. A split-screen inset shows the accelerometer reading: the moment the phone is released, the reading drops to zero and stays at zero throughout the fall. The phone strikes a foam pad below. At impact, the accelerometer reading spikes to over $50\ \text{m/s}^2$, oscillates wildly, then settles back to $9.8\ \text{m/s}^2$ as the phone comes to rest on the pad.]
Think about what you just saw. The phone is clearly accelerating --- it is falling under gravity at $a = g = 9.8\ \text{m/s}^2$. You can calculate this. You can see it in the video. Newton's second law demands it.
But the accelerometer reads zero.
This is not a malfunction. Every accelerometer on every phone in the world reads zero during free fall. Astronauts on the International Space Station --- who are perpetually in free fall --- live in a world where accelerometers read zero at all times. The instrument is working perfectly. It is just not measuring what you think it is measuring.
Before you read on: A phone accelerometer reads zero during free fall. Your textbook says the phone's acceleration is $g$ downward. Who is wrong --- the instrument or the textbook?
Write down your answer and your reasoning before continuing.
Neither is wrong. They are measuring different things. The textbook gives you the coordinate acceleration --- the second time derivative of position in an inertial reference frame. The accelerometer measures proper acceleration --- the acceleration relative to a local freely falling frame, which is equivalent to measuring the net non-gravitational force per unit mass.
When the phone sits on a table, the table pushes up on it with a normal force $N = mg$. The accelerometer detects this push and reports $g$ upward. When the phone is in free fall, the only force acting is gravity --- and gravity affects every atom of the phone and every atom of the accelerometer's internal sensor equally. There is nothing for the sensor to detect. It reads zero.
This distinction between coordinate acceleration and what instruments actually measure is not a footnote in physics. It is one of the central insights of general relativity, and it matters every time you try to connect a textbook equation to a real measurement.
This section is about that connection. Throughout this course, you have been building models of motion: projectiles, oscillators, collisions, rigid bodies. Those models produce predictions in the clean language of mathematics. But in the real world, predictions meet data --- and data comes from instruments, with all their limitations, conventions, and surprises. The skill you need now is not new physics. It is the skill of interpreting real measurements through the lens of the physics you already know.
Theory Meets Measurement
The accelerometer example reveals a general pattern. Here it is, stated plainly:
Every measurement is filtered through an instrument, and the instrument has its own physics.
An accelerometer does not measure $\ddot{x}$. A motion sensor does not measure position directly --- it measures the round-trip time of an ultrasonic pulse and converts it to distance. A video tracker gives you pixel coordinates that must be calibrated to physical coordinates. A force plate measures the total contact force, not the net force. A radar gun measures the component of velocity along the line of sight, not the full velocity vector.
None of these instruments are lying. But none of them hand you the textbook quantity directly. There is always a translation step, and that step requires understanding both the physics of the system and the physics of the measurement.
Pause and think: You use a motion sensor (ultrasonic range finder) to track a ball tossed straight up and caught. The sensor sits on the floor, pointing upward. Sketch what you expect the position-vs-time graph to look like. Now sketch the velocity-vs-time graph.
Here is the subtlety: the sensor measures distance from itself to the ball. When the ball is rising, the distance is increasing. When the ball is falling, the distance is decreasing. The velocity-vs-time graph is not a straight line with slope $-g$ --- it is a straight line with slope $-g$ only if you define "positive" as "upward." If the sensor reports distance (always positive), then the velocity graph shows a V-shape with a cusp at the top. Your interpretation depends on knowing how the sensor defines its output.
| Instrument | What it actually measures | Common assumption | The gap |
|---|---|---|---|
| Phone accelerometer | Non-gravitational force per unit mass (proper acceleration) | Coordinate acceleration $\ddot{x}$ | Reads zero in free fall; reads $g$ at rest |
| Ultrasonic motion sensor | Round-trip time of sound pulse, converted to distance | True position of the object | Measures distance from sensor, not from origin; can lose track if object tilts away |
| Video tracker | Pixel coordinates of a tracked feature | Physical coordinates in meters | Requires calibration; perspective distortion introduces error |
| Force plate | Total normal force on the plate surface | Net force on the object | Includes the object's weight; net force requires subtracting $mg$ |
| Radar/lidar speed sensor | Radial component of velocity (along line of sight) | Speed of the object | Underestimates speed if object moves at an angle to the sensor |
The table is not meant to be memorized. It is meant to illustrate a principle: the gap between what an instrument reports and what a textbook model predicts is not a flaw. It is a feature of doing physics in the real world. Closing that gap is a core professional skill for anyone who uses physics --- in research, in engineering, or in data science.
Building a Model and Confronting It with Data
Let's work through a concrete example, start to finish.
The setup
[Data set: A ball is dropped from rest and its vertical position is recorded every 1/30 of a second using video analysis. The data spans approximately 0.5 seconds of free fall, covering about 1.2 meters of drop. The data is presented as a two-column table: time $t$ (in seconds) and vertical position $y$ (in meters, with $y$ increasing downward from the release point).]
| $t$ (s) | $y$ (m) |
|---|---|
| 0.000 | 0.000 |
| 0.033 | 0.004 |
| 0.067 | 0.019 |
| 0.100 | 0.046 |
| 0.133 | 0.082 |
| 0.167 | 0.130 |
| 0.200 | 0.189 |
| 0.233 | 0.258 |
| 0.267 | 0.337 |
| 0.300 | 0.427 |
| 0.333 | 0.526 |
| 0.367 | 0.636 |
| 0.400 | 0.756 |
| 0.433 | 0.885 |
| 0.467 | 1.023 |
| 0.500 | 1.172 |
Step 1: Choose a model
For a ball dropped from rest in a uniform gravitational field with no air resistance, the textbook model is:
$$y(t) = \frac{1}{2}g t^2$$
This predicts a parabola with no linear term (since $v_0 = 0$) and no constant term (since $y_0 = 0$).
Step 2: Fit the model to the data
If the model is $y = \frac{1}{2}g t^2$, then fitting it means finding the value of $g$ that best matches the data. One approach: plot $y$ vs. $t^2$. If the model is correct, this should be a straight line through the origin with slope $g/2$.
[Interactive: Model Fitting Tool. The left panel shows a scatter plot of $y$ vs. $t$ with the data points above. The right panel shows $y$ vs. $t^2$ --- the same data, re-plotted. A slider lets the student adjust $g$ from 5 to 15 m/s$^2$. As $g$ changes, a parabola (left panel) and a straight line (right panel) update in real time. A readout shows the sum of squared residuals. The student adjusts $g$ to minimize the residuals.
Guided prompts: - "Start with $g = 9.8$. How well does the model fit?" - "Try $g = 9.4$. Is this better or worse?" - "Find the value of $g$ that minimizes the sum of squared residuals. Is it exactly $9.80$?"]
When you fit this data, you will find a best-fit value near $g \approx 9.4\ \text{m/s}^2$. Not $9.80$. The fit is decent --- the parabola tracks the data well --- but the extracted value of $g$ is about 4% low.
Pause and think: The best-fit $g$ is 4% below the accepted value. Before you dismiss this as "experimental error," consider: what specific physical effects could systematically reduce the apparent value of $g$?
List at least two possibilities.
Step 3: Examine the residuals
A residual is the difference between the measured value and the model prediction at each data point:
$$r_i = y_{\text{measured}}(t_i) - y_{\text{model}}(t_i)$$
If the model is perfect and the only errors are random measurement noise, the residuals should scatter randomly around zero with no pattern.
[Interactive: Residual Plot. Using the best-fit $g$, the tool plots the residuals $r_i$ vs. $t_i$. The student examines the pattern. If the residuals show a systematic trend (e.g., consistently negative at early times, positive at late times), the model is missing something.]
Look at the residuals carefully. If you see a systematic curve --- residuals that start near zero, drift negative, then become positive at later times --- that is not noise. That is a signature of missing physics.
Step 4: Identify the missing physics
The most likely cause of the 4%-low $g$ value is air resistance. As the ball accelerates, air drag increases (it is proportional to $v^2$ for a ball at moderate speeds). Drag opposes the downward motion, so the ball accelerates slightly less than $g$. Fitting a no-drag model to drag-affected data produces a best-fit $g$ that is systematically too low.
Another possibility: the ball was not perfectly at rest at $t = 0$. If it had a small upward velocity at release (from the hand opening), the early data would not match the $v_0 = 0$ assumption.
A third possibility: the video frame rate introduces a systematic timing offset. If the first recorded frame was not the exact instant of release, the time origin is slightly wrong.
Each of these is a different kind of deviation from the ideal model:
| Deviation | Type | What to do about it |
|---|---|---|
| Air resistance | Missing physics | Add drag to the model: $y = \frac{1}{2}g_{\text{eff}}t^2$ as an approximation, or solve the full drag equation |
| Nonzero initial velocity | Incorrect initial conditions | Fit a model with $v_0$ as a free parameter: $y = v_0 t + \frac{1}{2}g t^2$ |
| Timing offset | Instrumentation artifact | Fit with a time offset: $y = \frac{1}{2}g(t - t_0)^2$ |
| Random scatter in pixel positions | Measurement noise | Cannot be removed; accept it and quantify it with uncertainties |
This is the central skill of this section: looking at the gap between model and data and diagnosing what it means.
Three Categories of Deviation
When your model does not perfectly match your data, the mismatch falls into one of three categories. Learning to distinguish them is essential.
1. Noise
Random, uncorrelated scatter in the data. It comes from the finite precision of your instrument, vibrations, lighting changes in video analysis, electromagnetic interference in electronic sensors --- anything that adds random fluctuations to each measurement independently.
How to recognize it: Residuals that scatter randomly around zero with no pattern. If you repeat the experiment, the noise changes but the underlying trend does not.
What to do: You cannot eliminate noise from a single data set. You can reduce it by averaging multiple trials, by using a higher-precision instrument, or by fitting a model that smooths over the noise (a least-squares fit automatically does this).
2. Missing physics
A systematic deviation caused by a physical effect that your model does not include. Air resistance, friction, elastic deformation, thermal expansion, the finite mass of a pulley --- anything real that your idealized model neglected.
How to recognize it: Residuals that show a systematic pattern --- a trend, a curve, a growing offset. The pattern is reproducible. It gets worse (or changes character) when you change experimental conditions in a way that amplifies the missing effect (e.g., using a lighter ball increases the relative importance of air drag).
What to do: Improve the model. Add the missing effect and re-fit. If the residuals become random after adding the correction, you have likely found the right physics.
3. Instrumentation artifacts
Systematic deviations caused not by missing physics in the system but by the behavior of the instrument. A motion sensor that saturates at close range. A video tracker that introduces perspective distortion. An accelerometer with a calibration offset. A force plate with a resonance that rings after an impact.
How to recognize it: The deviation correlates with instrument-specific conditions rather than with the physics of the system. It may appear at particular positions (near the edge of a sensor's range), particular times (immediately after a sudden change), or particular frequencies (near the sensor's natural resonance).
What to do: Understand the instrument's limitations. Restrict your analysis to the range where the instrument is reliable. Apply instrument-specific corrections if they are known. Or switch to a different measurement technique.
Pause and think: An ultrasonic motion sensor tracks a cart bouncing between two spring bumpers on an air track. The position-vs-time data looks good in the middle of the track but shows strange spikes near each end. Is this noise, missing physics, or an instrumentation artifact?
Think about how an ultrasonic sensor works. It sends out a pulse and listens for the echo. If the cart is very close to the sensor, the echo returns before the sensor is ready to listen. If the cart tilts during a bounce, the echo may bounce away at an angle and miss the sensor entirely. These are instrumentation artifacts --- they tell you about the sensor, not about the cart.
Case Study: Phone Accelerometer During a Bounce
Let's return to the accelerometer and push the analysis further.
[Data set: A phone is dropped from 1.5 meters onto a hard rubber surface. The accelerometer records the vertical component of proper acceleration at 200 Hz (every 5 ms) for 2 seconds, starting 0.5 seconds before the drop. The data shows five distinct phases.]
Phase 1: At rest before the drop (t = 0 to 0.5 s)
The phone sits in a hand. The accelerometer reads $+9.8\ \text{m/s}^2$ (upward). This is the normal force from the hand, divided by the phone's mass. The phone is not accelerating in the coordinate sense, but the accelerometer reports $g$ because it measures proper acceleration.
Phase 2: Free fall (t = 0.5 s to ~1.05 s)
The phone is released. The reading drops to $0\ \text{m/s}^2$ and stays there for approximately 0.55 seconds. This is consistent with a fall from 1.5 m: the time to fall a distance $h$ from rest is $t = \sqrt{2h/g} = \sqrt{2(1.5)/9.8} \approx 0.55$ s.
Phase 3: First impact (t ~ 1.05 s)
A sharp spike, reaching $50$--$80\ g$ (depending on the surface and the phone), lasting about 5--10 ms. This is the contact force from the rubber surface decelerating the phone from its impact speed ($v = gt \approx 5.4\ \text{m/s}$) back to zero and then launching it upward.
Phase 4: First bounce (t ~ 1.06 to 1.15 s)
The phone is briefly airborne again. The accelerometer reads zero. This free-fall interval is shorter than the original fall because the bounce height is less than the drop height --- the collision is not perfectly elastic.
Phase 5: Subsequent bounces and settling
Repeated spikes (each smaller than the last) separated by shrinking intervals of zero readings. Eventually the phone comes to rest on the surface, and the reading settles back to $+9.8\ \text{m/s}^2$.
Pause and think: During Phase 3 (impact), the accelerometer spikes to $60g$. Does this mean the phone experiences a force of $60mg$?
Almost, but not exactly. The accelerometer reads proper acceleration, which is the non-gravitational force per unit mass. During impact, the non-gravitational force (the contact force from the surface) is indeed about $60mg$. But the net force on the phone is $60mg - mg = 59mg$ upward, giving a coordinate acceleration of $59g$ upward. The accelerometer reading and the coordinate acceleration differ by $g$ --- the same offset that caused the "zero in free fall" surprise.
Connecting data to a model
Can you model the bounce sequence? The free-fall phases are straightforward: $a = g$ downward, $v(t)$ linear, $y(t)$ quadratic. The duration of each free-fall phase tells you the bounce height, and the ratio of successive bounce heights gives you the coefficient of restitution $e$ (from Chapter 8):
$$e = \frac{v_{\text{after}}}{v_{\text{before}}} = \sqrt{\frac{h_{\text{bounce}}}{h_{\text{drop}}}}$$
The impact phases are harder to model because the contact force depends on the detailed mechanics of the collision --- material stiffness, deformation, energy dissipation. A simple model (constant deceleration during contact) predicts a rectangular pulse; the actual data shows a more complex shape. The gap between the simple model and the data reveals the real contact mechanics of the collision.
This is exactly the pattern: start with the simplest model, compare it to data, and let the residuals tell you what you are missing.
When Models Meet Engineering
The gap between theory and measurement is not just a laboratory curiosity. It is the central challenge of engineering.
Consider a structural engineer designing a pedestrian bridge. The model says:
- The bridge is a uniform beam supported at both ends.
- The maximum load before failure is determined by the yield strength of the steel.
- The predicted maximum load is 50,000 N distributed across the span.
The bridge is built. It is load-tested. It shows signs of stress at 42,000 N --- about 84% of the predicted maximum.
Is the model wrong? Is the steel defective? Is the construction flawed?
Pause and think: The bridge fails at 84% of the predicted load. List at least three possible reasons the model overpredicted. For each, identify whether the issue is missing physics, incorrect assumptions, or a gap between the idealized model and the real structure.
Here are some possibilities:
-
The beam is not perfectly uniform. Welds, bolt holes, and joint geometry create stress concentrations that the uniform-beam model ignores. These are locations where the local stress exceeds the average stress, causing failure earlier than the model predicts. (Missing physics: the model does not account for stress concentration.)
-
The load is not perfectly distributed. Real pedestrians cluster near the center. A concentrated load creates a larger maximum bending moment than the same total load distributed uniformly. (Incorrect assumption: the model assumed uniform loading.)
-
The material properties are not exactly as specified. Steel strength varies between batches. The yield strength used in the model is a nominal value; the actual steel may be slightly weaker. (Gap between model parameters and real parameters.)
-
Dynamic effects matter. Pedestrians walk, creating oscillating forces. If the walking frequency is near a natural frequency of the bridge, resonance amplifies the stresses beyond what a static load model predicts. (Missing physics: the model is static, but the real load is dynamic.)
-
Environmental degradation. Corrosion, temperature cycling, or fatigue from repeated loading may have weakened the structure before the load test. (Missing physics: the model assumes pristine material.)
Engineers handle this gap with safety factors --- they design structures to withstand loads 1.5 to 3 times larger than the expected maximum. The safety factor is an explicit acknowledgment that models are incomplete. It is not a sign of ignorance; it is a sign of wisdom. The engineer is saying: "My model captures the dominant physics, but I know it misses some things, and I am building in a margin for what I cannot predict."
This is the mature version of the skill you are developing in this section. In a course, you learn the ideal model first. In practice, you learn to estimate how far reality will deviate from the ideal --- and to build that estimate into your decisions.
A Framework for Interpreting Real Data
Here is a procedure you can follow whenever you need to connect a mechanics model to real measurements.
Step 1: State the model explicitly. Write down the equations, the assumptions, and the initial/boundary conditions. Be precise about what you are predicting.
Step 2: Understand your instrument. What does it actually measure? What is its range, resolution, and sampling rate? What are its known limitations?
Step 3: Translate between model and measurement. The model predicts $y(t)$; the instrument reports a voltage proportional to distance from the sensor. You need to know the conversion. This step is where many students lose track.
Step 4: Fit the model to the data. Use a least-squares fit, an eyeball fit, or whatever is appropriate. Extract the model parameters (like $g$ in the falling-ball example).
Step 5: Compute and examine the residuals. Plot them. Look for patterns. Are they random (noise) or systematic (missing physics or instrumentation artifacts)?
Step 6: Iterate. If the residuals reveal a systematic deviation, improve the model, re-fit, and check again. Keep going until the residuals are consistent with noise alone --- or until you have identified why they are not.
This is the scientific method in miniature. It is also the engineering design cycle, the data science pipeline, and the debugging process. The specific physics changes; the structure does not.
Practice
Layer 1: Concrete
Problem 1. A ball is dropped from rest and video-tracked. The following data gives position $y$ (downward) vs. time $t$. Fit the model $y = \frac{1}{2}g t^2$ to the data using a $y$ vs. $t^2$ plot, extract a value for $g$, and compute the residuals.
| $t$ (s) | $y$ (m) |
|---|---|
| 0.00 | 0.000 |
| 0.05 | 0.013 |
| 0.10 | 0.048 |
| 0.15 | 0.107 |
| 0.20 | 0.190 |
| 0.25 | 0.295 |
| 0.30 | 0.425 |
| 0.35 | 0.576 |
| 0.40 | 0.750 |
Check your answer
Rewrite the data as $y$ vs. $t^2$. If $y = \frac{1}{2}g t^2$, then plotting $y$ against $t^2$ should give a straight line through the origin with slope $g/2$. | $t^2$ (s$^2$) | $y$ (m) | |:---|:---| | 0.0000 | 0.000 | | 0.0025 | 0.013 | | 0.0100 | 0.048 | | 0.0225 | 0.107 | | 0.0400 | 0.190 | | 0.0625 | 0.295 | | 0.0900 | 0.425 | | 0.1225 | 0.576 | | 0.1600 | 0.750 | The best-fit slope through the origin is approximately $g/2 \approx 4.69\ \text{m/s}^2$, giving $g \approx 9.38\ \text{m/s}^2$. (You can compute this by least-squares: slope $= \sum y_i t_i^2 / \sum t_i^4$.) The residuals are $r_i = y_i - \frac{1}{2}(9.38)t_i^2$. For the first few points, the residuals are very small and positive. For later points, the residuals grow slightly negative, suggesting the ball is falling a bit slower than the constant-$g$ model predicts. This is consistent with air drag becoming more significant as the ball speeds up.Problem 2. A phone accelerometer records vertical proper acceleration during a drop-and-bounce experiment. During free fall, the reading is $0.3\ \text{m/s}^2$ (not exactly zero). During rest on the table afterward, the reading is $9.6\ \text{m/s}^2$ (not exactly $9.8$). What do these offsets tell you, and how would you correct for them?
Check your answer
The readings suggest a **calibration offset**. The accelerometer has a small systematic bias. During free fall, the true proper acceleration is $0\ \text{m/s}^2$, but the instrument reads $0.3\ \text{m/s}^2$. During rest on the table, the true proper acceleration is $g = 9.8\ \text{m/s}^2$, but it reads $9.6\ \text{m/s}^2$. You can model this as a linear calibration: $a_{\text{true}} = \alpha \cdot a_{\text{reading}} + \beta$. Using the two known points: - Free fall: $0 = \alpha(0.3) + \beta$ - At rest: $9.8 = \alpha(9.6) + \beta$ Solving: from the first equation, $\beta = -0.3\alpha$. Substituting into the second: $9.8 = 9.6\alpha - 0.3\alpha = 9.3\alpha$, so $\alpha \approx 1.054$. Then $\beta \approx -0.316$. The corrected acceleration is $a_{\text{true}} \approx 1.054 \cdot a_{\text{reading}} - 0.316$. This is an instrumentation artifact, not missing physics. The phone's sensor has a slight gain error and a slight offset. Once corrected, the data can be used reliably.Layer 2: Pattern
Problem 3. For each of the following scenarios, identify whether the deviation between the ideal model and the data is most likely due to (a) noise, (b) missing physics, or (c) an instrumentation artifact. Justify your classification and propose what you would do about it.
(i) A pendulum is timed for 50 swings. The period is measured as $2.014 \pm 0.003$ s. The model $T = 2\pi\sqrt{L/g}$ predicts $T = 2.006$ s. The discrepancy is consistent across repeated trials.
(ii) A motion sensor tracks a cart on an air track. The velocity-vs-time data shows a linear decrease (constant deceleration) as expected from friction, but every 0.2 seconds there is a brief upward spike in the velocity data.
(iii) A spring-mass system is set oscillating and its position is recorded for 30 seconds. The amplitude at the beginning and end of the 30 seconds appears the same, but individual data points scatter around the fitted sinusoidal curve by about 2 mm.
Check your answer
**(i) Missing physics.** The discrepancy is systematic and reproducible, not random. The simple pendulum model $T = 2\pi\sqrt{L/g}$ assumes small angles. If the pendulum swings at a moderate amplitude (say, 15--20 degrees), the true period is slightly longer than the small-angle prediction. The correction for finite amplitude is $T \approx 2\pi\sqrt{L/g}\left(1 + \frac{1}{16}\theta_0^2 + \cdots\right)$. To fix: either reduce the amplitude to make the small-angle approximation more accurate, or use the corrected formula. **(ii) Instrumentation artifact.** The spikes occur at regular intervals unrelated to the cart's physics, suggesting they are caused by the sensor --- perhaps the ultrasonic pulse occasionally reflects off a nearby surface (multipath interference) or the sensor briefly loses tracking. To fix: filter out the spikes or restrict the analysis to the smooth portions of the data. **(iii) Noise.** The 2 mm scatter is random and does not grow or show a pattern over time. The underlying sinusoidal model fits well (no amplitude decay, meaning damping is negligible on this timescale). The scatter is likely due to the finite resolution of the position sensor. To fix: accept the noise and report uncertainties, or average over multiple cycles.Layer 3: Structure
Problem 4. "What is the difference between 'noise' and 'missing physics'? How do you tell?"
Check your answer
**Noise** is random and uncorrelated. It has no systematic pattern. If you repeat the experiment, the noise is different each time, but the underlying signal is the same. Noise does not depend on the physical parameters of the system in any systematic way. In a residual plot, noise looks like random scatter. **Missing physics** is systematic and reproducible. It creates a pattern in the residuals --- a trend, a curve, a growing or oscillating deviation. If you repeat the experiment under the same conditions, the same pattern appears. Crucially, missing physics responds to changes in experimental conditions: if you change a parameter that affects the missing effect (e.g., switch to a heavier ball, which makes air drag relatively less important), the pattern changes in a predictable way. The key test: **change the experimental conditions and see if the deviation changes systematically.** - If you drop a heavier ball and the best-fit $g$ gets closer to $9.8$, the deviation was caused by air drag (missing physics). - If you drop a heavier ball and the scatter stays the same size, the scatter is noise. - If you move the motion sensor to a different position and the spikes disappear, the spikes were instrumentation artifacts. This is why experimentalists repeat measurements under varied conditions, not just under identical conditions. Identical repetitions reveal noise. Varied conditions reveal missing physics.Layer 4: Transfer
Problem 5. An engineer's finite-element model of a bridge predicts a maximum load of 100 kN before structural failure. The actual bridge, when load-tested, shows the first signs of yielding at 85 kN. The engineer's model assumed:
- The steel has a yield strength of 250 MPa (the nominal specification).
- The beam has a uniform cross-section.
- All connections are perfectly rigid.
- The load is applied slowly (quasi-static).
What might the model have missed? For each possibility you identify, state whether it would cause the model to overpredict or underpredict the failure load, and explain why.
Check your answer
Several factors could account for the 15% overprediction: 1. **Stress concentrations at joints.** Welded or bolted connections create geometric discontinuities where local stress exceeds the average stress by a factor of 2 to 5. The uniform-beam model ignores these. *Overpredicts* the failure load because it underestimates the peak stress. 2. **Material variability.** The actual yield strength may be lower than the nominal 250 MPa. Steel properties vary between heats and are reported as minimum guaranteed values, but individual specimens can fall below. *Overpredicts* if the actual material is weaker than assumed. 3. **Non-rigid connections.** If bolted connections allow slight rotation or slippage, the beam does not behave as a single rigid element. This can change the distribution of bending moments and create unexpected stress patterns. Could *overpredicts* or *underpredict*, depending on the geometry. 4. **Residual stresses from fabrication.** Welding and cold-forming introduce residual stresses in the steel that add to the applied stresses. The model does not include these. *Overpredicts* because the steel is already partially stressed before any external load is applied. 5. **Dynamic loading effects.** If the "slow" load test is not slow enough, or if the structure has a vibration mode excited by the loading process, dynamic amplification increases the effective load. *Overpredicts* the static failure load relative to the dynamic reality. 6. **Imperfect geometry.** The as-built beam may not have a perfectly uniform cross-section. Slight misalignment, out-of-straightness, or thickness variation can reduce the actual load-carrying capacity. *Overpredicts.* Notice the pattern: almost every simplification in the model acts in the same direction --- it makes the structure appear stronger than it is. This is why safety factors are not optional luxuries but engineering necessities. The model captures the dominant physics (beam bending, material yielding) but systematically misses the details that weaken the real structure. The 15% gap in this problem is typical of what engineers encounter in practice. A well-calibrated model for a well-understood structure might overpredicate by 10--20%. For novel structures or extreme conditions, the gap can be much larger.Reflection
Think back to the beginning of this course. In Section 2.6, you first encountered the idea of fitting a kinematic model to data and looking at residuals. At that point, the models were simple --- constant velocity, constant acceleration --- and the data was clean.
Now you have the full range of mechanics at your disposal: forces, energy, momentum, rotational dynamics, oscillations. And you have seen that real data is messy, instruments have their own physics, and models always leave something out.
How has your understanding of "real" vs. "ideal" physics changed over this course?
Consider these questions:
- When you see a discrepancy between a model and a measurement, is your first instinct still to assume you made an error? Or do you now think more carefully about what the discrepancy might be telling you?
- Do you view the simplifying assumptions in a model (frictionless surfaces, massless ropes, point particles) as defects, or as deliberate choices that make the problem tractable while capturing the essential physics?
- An engineer builds a safety factor into a design. A physicist adds a perturbation correction to a model. A data scientist adds a regularization term to a fit. Are these the same idea in different languages?
The gap between theory and reality is not a failure of physics. It is where physics lives. Every real advance --- in science, in engineering, in understanding --- comes from taking that gap seriously, diagnosing its causes, and deciding what to do about it.
Looking Ahead
You have spent this section learning to read the gap between models and measurements --- to distinguish noise from missing physics from instrumentation artifacts, and to iterate your model until the residuals tell you nothing more.
In the next section, you will apply this skill (along with everything else you have learned in this course) to rich, multi-concept applications: vehicles braking and turning, athletes in motion, machines transmitting force and torque. These are systems where no single chapter of this course suffices. You will need kinematics, forces, energy, momentum, and rotational dynamics simultaneously --- and you will need the modeling judgment you developed here to decide which effects matter and which can be safely ignored. Section 15.5 is where the full toolkit meets the real world.