🧬 biology

Why your ODE fit fails, and five steps to fix it: a practical guide

This paper presents a practical five-step recipe—featuring problem rescaling, tightened solver tolerances, gradient-matching initialization, residual analysis, and log-scale interval reporting—to overcome common failures in fitting ODE models to data, demonstrating that these steps significantly improve success rates and reveal model flaws in both synthetic and real-world biological examples.

Original authors: Mohammad Jamhuri, Putri Wulan Agustina

Published 2026-10-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Jamhuri, Putri Wulan Agustina

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the world of mathematical biology, scientists build virtual models to understand how living systems change over time. These models are essentially sets of rules describing how populations grow, how diseases spread, or how drugs move through the body. To make these virtual worlds match reality, researchers must find the specific numbers that govern the rules, such as how fast a virus spreads or how quickly a predator eats. This process, known as fitting a model to data, is a routine task in the field. However, it is deceptively difficult. The computer programs used to find these numbers often stop too early, convinced they have found the best answer when they have actually settled for a poor approximation. When this happens, the resulting predictions can be wildly inaccurate, leading scientists to draw wrong conclusions about the very systems they are trying to understand.

A new guide by researchers Mohammad Jamhuri and Putri Wulan Agustina investigates why these computer fits fail so often and offers a straightforward, five-step method to fix them. The authors discovered that the most common cause of failure is not a flaw in the model or the data, but a hidden setting in the software itself. When scientists use standard tools to solve the equations, the software defaults to a setting that allows for a small amount of error in every calculation. While this error seems tiny, it accumulates rapidly in complex, changing systems, causing the computer to calculate the direction of improvement incorrectly. The researchers found that simply tightening this setting to be much more precise dramatically increases the chance of finding the correct answer, turning a process that often fails into one that succeeds almost every time.

The researchers tested their method on several classic problems, including the famous predator-prey relationship between hares and lynx, and the spread of influenza in a boarding school. In one striking test involving the predator-prey model, the standard software settings led to a successful fit in only about one-third of the attempts. When the researchers tightened the precision setting, the success rate jumped to nearly two-thirds, and the estimated numbers became far more accurate. They showed that this improvement happened regardless of which computer program or mathematical strategy was used to solve the equations. The key insight is that the error comes from the way the software calculates the slope of the solution, and this calculation is only as good as the precision of the underlying numbers. By demanding higher precision, the researchers removed the noise that was confusing the computer.

The guide does not stop at just fixing the numbers; it provides a complete workflow to ensure the model itself is trustworthy. The first step involves rescaling the numbers so that very large and very small values can be compared fairly. The second step, which the authors identify as the most critical, is to tighten the solver tolerance to a specific, high-precision level. The third step ensures the computer starts its search from a good guess, using a technique that looks at the rate of change in the data to find a starting point close to the truth. The fourth step is a reality check: plotting the differences between the model and the actual data to see if they look like random noise or if they follow a pattern. If the differences show a pattern, it means the model is missing a piece of the puzzle. The final step involves calculating the range of uncertainty for the results, ensuring that the reported numbers reflect the true limits of what the data can tell us.

Applying this full five-step process to real-world data revealed hidden flaws in established models. When the researchers fitted a standard disease model to the 1978 influenza outbreak in an English boarding school, the fit looked good at first glance. However, the residual check—the fourth step—showed a clear pattern in the errors, indicating that the simple model could not capture the true speed of the recovery. By adjusting the model to account for a more realistic recovery process, they improved the fit and changed the estimate of how contagious the virus was, lowering the basic reproduction number from nearly four to just over three. This shift was significant enough to alter the understanding of the outbreak's severity, proving that a better numerical fit can lead to better biological insights.

The authors emphasize that these steps are not just for experts with specialized software; they are practical habits for anyone using general-purpose tools. The most surprising finding is that the default settings in popular software packages, which are used by thousands of researchers, are often too loose for reliable modeling. By adopting these five habits—rescaling, tightening precision, starting with a good guess, checking the errors, and reporting honest uncertainty ranges—scientists can avoid the trap of false confidence. The guide concludes that while the mathematics of these models can be complex, the path to a reliable answer is often just a matter of paying attention to the details of how the computer calculates the solution.

Technical Summary: "Why your ODE fit fails, and five steps to fix it: a practical guide"

Problem Statement
Fitting ordinary differential equation (ODE) models to data is a routine task in mathematical biology, yet initial attempts frequently yield poor fits or unstable parameter estimates that vary with the initial guess. The authors identify that these failures rarely announce themselves; optimizers often report "success" despite returning parameters that are significantly incorrect. The paper argues that approaches using general-purpose solvers and least-squares routines with default settings mask three distinct difficulties:

Scale Disparities: Parameters often differ by orders of magnitude.
Inexact Jacobians: The finite-difference Jacobian, required by most optimizers, inherits errors from the numerical ODE solution.
Non-Convexity: The cost surface often contains multiple local minima, particularly in oscillating systems.

The central finding of the paper is that the default relative solver tolerance (typically $10^{-3}$ in Python, MATLAB, and Julia) is a primary, often invisible, cause of failure. At this default setting, the finite-difference Jacobian can be up to 31% inaccurate due to the "internal differentiation effect," where the solver's discretization error is differentiated along with the solution.

Methodology: The Five-Step Recipe
The authors propose a five-step recipe applicable to any ODE solver and least-squares routine. Each step includes a specific numerical setting and a "warning sign" indicating failure if the step is skipped.

Rescale Parameters and Residuals:
- Action: Estimate parameters on the log scale ( $\theta = \log p$ ) to keep them positive and normalize step sizes. Scale residuals to match the noise model (e.g., using log-differences for proportional noise or square-root differences for Poisson-like counts).
- Warning Sign: Estimates hitting zero/negative values or one output being fitted well while others are ignored.
Tighten Solver Tolerance:
- Action: Set the relative solver tolerance ($rtol$) to $10^{-6}$ and the absolute tolerance ($atol$) three orders of magnitude below the smallest state of interest.
- Rationale: This reduces the error in the finite-difference Jacobian from ~31% (at $10^{-3}$ ) to <0.2%. The paper notes that tuning the finite-difference step size is less effective than tightening the solver tolerance.
- Warning Sign: The fit stops early with a "success" message, or estimates change significantly when the tolerance is altered.
Start Close to the Answer:
- Action: Use gradient matching (fitting a spline to data and differentiating to estimate derivatives) if all states are observed. Otherwise, use a multistart approach (10 initial guesses spread over a plausible log-scale range).
- Warning Sign: Different initial guesses yield different optima with Sum of Squared Errors (SSE) differing by more than 1%.
Check Residuals:
- Action: Plot scaled residuals against time. Verify they resemble noise (no trends/waves) by checking that lag-1 autocorrelation is below $2/\sqrt{n}$ and the number of runs is near $n/2 + 1$ .
- Warning Sign: Residuals show a systematic pattern (waves or trends) or high autocorrelation, indicating a structural model flaw.
Report Intervals on the Log Scale:
- Action: Calculate confidence intervals on the log scale using the Jacobian covariance matrix. For parameters with wide intervals, verify using profile likelihood to check for non-quadratic behavior or identifiability issues.
- Warning Sign: A profile likelihood that is flat or open on one side, or an interval spanning more than an order of magnitude.

Key Results
The authors validated the recipe using four textbook models (Logistic growth, SIR, two-compartment Pharmacokinetic, and Lotka–Volterra) and two classic datasets (1978 boarding school influenza and Hudson Bay lynx–hare data).

Solver Tolerance Impact: On the Lotka–Volterra model, tightening the tolerance from $10^{-3}$ to $10^{-6}$ increased the success rate of fits (within 5% of the global optimum) from 30–36% to 58–68% across three datasets and two optimizers.
Ablation Study: In 400 trials across four models, no single step sufficed for all cases. However, applying the full recipe achieved the best fit in 100% of trials.
- Rescaling (Step 1) was most critical for the Pharmacokinetic model.
- Tightening tolerance (Step 2) was essential for oscillating models (Lotka–Volterra) and the SIR model.
- Gradient matching (Step 3) provided the most efficient start for the fully observed Lotka–Volterra model, reaching the global optimum in a single run with fewer evaluations than random starts.
Real-World Application (Influenza Data): Applying Step 4 (residual check) to the 1978 influenza data revealed a systematic pattern in the residuals of the standard SIR model (lag-1 autocorrelation of 0.66 vs. threshold 0.53). This indicated the model missed a mechanism. The authors refined the model to a "two-stage" SIR (splitting the infected state), which removed the residual pattern, lowered the autocorrelation to 0.34, and revised the basic reproduction number ( $R_0$ ) estimate from 3.96 to 3.18.

Significance and Claims
The paper positions itself as a practical guide for newcomers fitting small models with general-purpose tools, complementing existing literature on specialized software and statistical theory. Its primary contributions are:

Quantifying the "Hidden" Failure: Demonstrating that the default solver tolerance is the main cause of failed fits in standard environments, not the optimizer or the difference step.
Actionable Protocol: Providing a measured, step-by-step workflow where each step addresses a specific failure mode, validated by empirical success rates.
Model Diagnosis: Showing that residual checks (Step 4) can serve as a diagnostic tool to identify structural model flaws (e.g., non-exponential residence times) rather than just assessing fit quality.

The authors conclude that while the individual techniques are established, their measured effect and the specific combination of these five steps provide a reliable, low-cost method to ensure robust ODE model calibration. The guide emphasizes that for oscillating systems, the default $10^{-3}$ tolerance roughly halves the success rate of standard explicit solvers, making the proposed $10^{-6}$ tolerance a critical, often overlooked, adjustment.

Technical Summary: "Why your ODE fit fails, and five steps to fix it: a practical guide"

More like this