1 — What this tab does

Sigmoidal growth-curve fitting for closed batch fermentations. From time-series biomass data (and optionally substrate and product concentrations), the calculator extracts kinetic descriptors: maximum specific growth rate μ_max, lag time λ, carrying capacity X_max, doubling time t_d, and exponential / stationary phase durations — plus, when substrate/product columns are present, yield coefficients Y_X/S, Y_P/S and specific rates q_S, q_P.

Use this tab when: you have a single closed run with no feed addition, growth follows a roughly sigmoidal pattern (lag → exponential → stationary), and you want one specific growth rate that summarises the run.

Use a different tab when:

• Substrate is added during the run → Tab 2 (Fed-Batch)
• You have multiple steady-state operating points at different dilution rates → Tab 3 (Continuous)
• The culture switches substrates mid-run (e.g. glucose → ethanol) → Tab 4 (Diauxic)

Forward simulation. If you also need to predict a fermentation trajectory from chosen kinetic parameters — rather than fit a measured one — the Simulation sub-tab on Tab 2 (Fed-Batch) integrates the underlying ODE system forward under user-chosen feed strategies and initial conditions. Tab 1 (this tab) is closed-batch only.

2 — Theory & equations

All five available models are empirical sigmoidal forms. They differ in how lag, inflection asymmetry, and shape parameters are handled.

Logistic (3 params): symmetric S-curve, no lag.

Modified Gompertz (Zwietering 1990) (4 params): asymmetric S-curve with an explicit lag term — the default for batch fermentation.

Baranyi-Roberts (1994) (4 params): physiologically motivated lag with a sharp lag-to-exponential transition. Preferred for low-inoculum or stress-adapted cultures.

Richards (1959) (4 params): generalised logistic with a tunable shape parameter ν controlling the inflection asymmetry.

Gompertz (basic) (3 params): asymmetric S-curve with early inflection but no explicit lag.

All models are fitted by Levenberg-Marquardt nonlinear regression with a numerical Jacobian. Initial guesses come from data heuristics: X₀ from the first point, X_max from the 90th percentile, μ_max from the steepest consecutive log-slope, λ from the first time biomass exceeds X₀ by >5%.

3 — Data format

Each row is one timepoint. Columns:

• Required: time, biomass
• Optional: substrate, product

Whitespace, tab, comma, or semicolon separators all work. The first row is auto-detected as a header (any non-numeric values) and skipped.

Biomass can be in any consistent unit within a dataset (g DCW/L, OD₆₀₀, cell count) — μ_max and λ are unit-independent. Substrate and product are typically g/L for yields to be dimensionally meaningful.

Minimum 4 rows for a 3-parameter model; 6+ rows recommended for 4-parameter models. Do not include post-decline data — the models fit growth + stationary only; a death phase will distort X_max.

4 — Strain library

The Strain dropdown above each data box gives 7 preset datasets, each ODE-validated against canonical yeast literature:

• S. cerevisiae S288c — haploid laboratory reference (Saccharomyces Genome Database)
• S. cerevisiae CEN.PK113-7D — metabolic-engineering benchmark strain
• Ethanol Red (Lesaffre) — first-generation industrial fuel ethanol
• PE-2 (Brazilian sugarcane) — Brazilian cane-ethanol industrial workhorse
• Thermosacc (Lallemand) — thermotolerant fuel ethanol
• WLP001 (California Ale) — brewing standard
• EC-1118 (Lalvin wine) — wine fermentation

Selecting a strain replaces the data textarea with that strain's preset time-series. To use your own data, choose Custom strain — the dropdown is replaced by a text input where you can name your strain, and the data textarea is cleared for pasting.

Defaults: Reference = S288c, Novel = Ethanol Red. These two cover the laboratory-vs-industrial contrast that motivates most batch-screening experiments.

5 — Models & fits

• Logistic — use when data shows a smooth approach to asymptote and lag is short or absent. Three parameters: X₀, X_max, μ_max.
• Gompertz (basic) — use when growth approaches the asymptote slowly; the early-inflection asymmetry helps when the upper plateau is well-sampled. Three parameters: X_max, b, c.
• Modified Gompertz (Zwietering) — default. Asymmetric S-curve with explicit lag. Recommended unless you have a specific reason to choose otherwise. Four parameters: X₀, A, μ_max, λ.
• Baranyi-Roberts — preferred when the lag-to-exponential transition is sharp (low-inoculum, stress-adapted, or pre-incubated cells). Four parameters: X₀, X_max, μ_max, λ.
• Richards — flexible but often over-parameterised; use only when inflection asymmetry needs explicit tuning. Four parameters: X_max, μ_max, τ, ν.

Pick the model with the lowest AIC — it penalises extra parameters, so a 4-param model only "wins" if it explains the data meaningfully better than a 3-param model. Auto-select best automates this and shows a ranked comparison table.

6 — Workflow

1. Pick a Reference strain from the left card's dropdown, or choose Custom strain and paste your own data.
2. Pick a Novel strain from the right card, or paste custom data.
3. Optionally edit the strain identity fields (name, substrate, major/minor product). These are documentation-only; not used in any calculation.
4. Pick a growth model. Default is Modified Gompertz.
5. Click Run analysis for the selected model, or Auto-select best to fit all 5 models and pick the lowest AIC.
6. Inspect the parameter table and the fitted-curve plot. Look for systematic residual structure on the plot — a curve in the residuals indicates the model is misspecified for this data.
7. Set the Metabolism dropdown to override the inferred regime, or leave on Auto.
8. Repeat for the other strain. Once both are fit, the bottom Reference vs Novel comparison table populates with side-by-side parameters and Δ Novel/Ref percent differences.

7 — Hybrid metabolism classifier

After a fit completes — provided substrate and product columns are in the data — the calculator computes Y_X/S (biomass yield on substrate) and Y_P/S (product yield on substrate) and infers the metabolic regime:

• Respiratory: Y_X/S ≥ 0.45 g/g, Y_P/S < 0.05 g/g — fully aerobic, no overflow metabolism.
• Respiro-fermentative: 0.20 < Y_X/S < 0.45 g/g — Crabtree-positive aerobic with partial overflow to ethanol.
• Fermentative: Y_X/S < 0.20 g/g, Y_P/S 0.40–0.50 g/g — anaerobic or fully fermentative.

The dropdown lets you override (Auto / Respiratory / Respiro-fermentative / Fermentative). If the regime you select conflicts with the inferred one, a ⚠ Mismatch flag appears in the parameter table. This is a useful sanity check: if you ran an aerobic batch and the classifier reports "Fermentative," either oxygen transfer was limiting or your Y_X/S measurement is off.

8 — Reference vs Novel comparison

Once both strains have been fit, the bottom summary table reports for each parameter the Reference value, the Novel value, and Δ Novel/Ref — the percent change relative to Reference.

Use this to quantify whether your engineered strain has the property you intended. A +20% gain in μ_max is meaningful; a +2% gain is within typical batch-to-batch noise.

9 — Output interpretation

The Extracted Parameters table is grouped:

• Model parameters — raw parameters returned by the fit (e.g. X₀, A, μ_max, λ for Modified Gompertz).
• Derived kinetic outputs — μ_max, t_d = ln(2)/μ_max, X_max, λ (when the model has a lag term), exponential and stationary phase durations.
• Yields & specific rates — Y_X/S, Y_P/S, q_S, q_P. Only present when substrate/product columns are in the data.
• Metabolism — user choice + inferred regime + mismatch flag if they conflict.
• Fit quality — R², RMSE, AIC, n / k.

The plot overlays measured points on the fitted curve. R² > 0.99 is expected for a clean batch; R² 0.95–0.99 is acceptable; < 0.95 indicates either noisy data or the wrong model.

10 — Worked example

Default Reference: S288c on glucose, aerobic-fermentative regime (loaded by default in the Reference card).

• Initial: X₀ = 0.10 g/L, S₀ = 30 g/L glucose
• After 24 h: X ≈ 3.86 g/L, S ≈ 0 g/L, P (ethanol) ≈ 1.5 g/L

Modified Gompertz fit on this dataset typically returns:

• μ_max ≈ 0.42 h⁻¹, t_d ≈ 1.65 h, λ ≈ 4 h, X_max ≈ 3.85 g/L, R² > 0.999
• Y_X/S ≈ (3.86 − 0.10) / (30 − 0) = 0.125 g X/g S — below the 0.20 cut-off, so the classifier reports Fermentative.
• Y_P/S ≈ 1.5 / 30 = 0.05 g P/g S — low ethanol yield suggests significant respiration despite the low Y_X/S (mixed metabolism, consistent with Crabtree-positive aerobic operation).

The default Novel is Ethanol Red. Running both should show Ethanol Red with higher μ_max and shorter t_d, reflecting its industrial selection for fermentation rate. The comparison table flags these gains in the Δ column.

11 — Edge cases & troubleshooting

• Fit fails to converge: try Auto-select best. If all 5 models fail, your data is too sparse, contains post-decline points, or starts mid-exponential. Strip rows after X_max is reached.
• R² < 0.95: visually inspect the plot. Systematic curve in residuals → wrong model; random scatter → noisy data, parameters have wide uncertainty.
• λ pinned at 0: your data starts at or after the inflection point. The model can't recover lag information from later points alone. Either include earlier data or use Logistic (no lag term).
• X_max exactly equals last data point: data didn't reach stationary phase; X_max is unreliable. Extend the run, or read μ_max only.
• Negative or absurd parameter values: the algorithm hit a local optimum. Try a different model first.
• No yield outputs: confirm your data has 3+ columns. Yields require substrate values; product yields require product values.
• Auto-select best picks Richards on noisy data: Richards has 4 free parameters and can over-fit noise. Cross-check by running Modified Gompertz manually — if its AIC is within ~2 of Richards, prefer the simpler model.
• "Mismatch" flag in metabolism: your stated regime conflicts with the Y_X/S/Y_P/S-inferred regime. Check oxygen-transfer rate, glucose feed rate, or your yield measurement.

12 — Acknowledged limitations

• Fitted parameters are batch-curve descriptors, not fundamental kinetic constants. K_S, K_I, true Y_X/S, and maintenance m_S require substrate-varied data — use Tab 3 (Continuous) for those.
• This calculator does not estimate parameter confidence intervals. For standard errors, use R nls or Python scipy.optimize.curve_fit which return the covariance matrix.
• Levenberg-Marquardt uses heuristic initial guesses; very noisy or atypically shaped datasets may converge to local optima. If fit quality is poor, try a different model first, then check data quality.
• Yields are batch-integral estimates over the full run. Specific rates q_S and q_P are reported as windowed averages over the exponential phase. These do not replace continuous-culture chemostat measurements for parameter precision.

1 — Why batch growth follows a sigmoidal curve

In a closed batch, all substrate is loaded at the start. Cells progress through four characteristic phases — lag (metabolic adaptation, no division), exponential (μ ≈ μ_max, substrate non-limiting), deceleration (substrate falling, μ declining), and stationary (substrate exhausted or product-inhibited, μ → 0). The biomass trajectory X(t) plotted against time traces an asymmetric S-curve: a slow start, an inflection where dX/dt is maximal, and a slow approach to an asymptote X_max.

The differential form is simple:

dX/dt = μ(X, S) · X

but μ depends on substrate concentration through Monod-style saturation, on biomass through density-dependent feedback (waste accumulation, oxygen depletion, ethanol toxicity), and on time through the lag-phase adaptation. Closed-form integration of the full mechanistic system is intractable except in trivial cases. The practical approach is to fit X(t) directly with empirical sigmoidal functions whose parameters happen to coincide with biologically meaningful quantities.

2 — Five sigmoidal models

The calculator implements five canonical models. They differ in symmetry around the inflection point, in how lag time is parameterized, and in their parameter count (3 vs 4). All five fit batch data through phenomenological matching rather than mechanistic derivation, but each emerges from a defensible theoretical limit.

2.1 Logistic (Verhulst 1838)

The earliest sigmoidal growth law — originally proposed for human population growth — assumes the per-capita growth rate declines linearly with population size:

dX/dt = μ_max · X · (1 − X/X_max)

which integrates to:

X(t) = X_max / (1 + (X_max/X₀ − 1) · exp(−μ_max t))

Three parameters: X₀, X_max, μ_max. The curve is symmetric around the inflection point at X = X_max/2; there is no explicit lag term. Use when lag is short or absent and the data is close to symmetric.

2.2 Gompertz (Gompertz 1825)

Originally proposed for human-mortality probabilities, Gompertz curves are asymmetric: the inflection occurs early, at X = X_max/e ≈ 0.37 X_max, and the approach to the asymptote is slow.

X(t) = X_max · exp(−b · exp(−c t))

Three parameters: X_max, b, c. The maximum specific growth rate is μ_max = b · c / e (rate at the inflection point divided by biomass at the inflection point). Use when growth approaches the asymptote slowly with no lag.

2.3 Modified Gompertz / Zwietering reparameterization (Zwietering 1990)

The default model. Zwietering showed that the basic Gompertz function can be rewritten so that its parameters are directly the biologically meaningful quantities — lag time, maximum specific growth rate, and asymptotic biomass — rather than the abstract b and c:

X(t) = X₀ + A · exp(−exp((μ_max · e / A)(λ − t) + 1))

Four parameters: X₀, A = X_max − X₀, μ_max, λ (lag time). Using A as the additive amplitude (rather than X_max directly) keeps the parameters orthogonal — small changes in μ_max and λ don't bleed into A. This is the recommended default for batch fermentation: the lag is explicit, the curve is asymmetric, and the parameters are interpretable.

2.4 Baranyi-Roberts (Baranyi & Roberts 1994)

A dynamic model with mechanistic motivation. The lag is treated as a quantity of an unspecified intracellular Q(t) (sometimes interpreted as substrate adaptation, RNA pool, or membrane fluidity) that must accumulate before growth begins. The full expression is:

ln(X/X₀) = μ_max · A(t) − ln(1 + (e^{μ_max A(t)} − 1) / (X_max/X₀))

where A(t) = t + (1/μ_max) · ln(e^{−μ_max t} + e^−h₀ − e^{−μ_max t − h₀}) and h₀ = μ_max · λ.

Four parameters: X₀, X_max, μ_max, λ. Compared to Modified Gompertz, the lag-to-exponential transition is sharper. Recommended for low-inoculum or stress-adapted cultures where the entry into exponential growth is abrupt.

2.5 Richards (Richards 1959)

An empirical generalization of the logistic with an extra shape parameter ν that lets the inflection point move:

X(t) = X_max · (1 + ν · exp(−μ_max(t − τ)))^−1/ν

Four parameters: X_max, μ_max, τ (time of inflection), ν (shape). When ν → 0 the curve approaches Gompertz; when ν = 1 it equals the Logistic; intermediate values give a flexible asymmetric shape. There is no explicit lag term, so τ absorbs both the lag and the position of the inflection. Use when the data shows clear asymmetry that doesn't match canonical Gompertz, or as a tiebreaker when no other model fits well.

3 — Nonlinear regression: Levenberg-Marquardt

Fitting any of these models to data X_i at times t_i means solving an iterative least-squares problem. Define the residuals:

r_i(p) = X_i − f(t_i, p)

where p is the parameter vector and f is the chosen growth model. The objective is:

SSE(p) = Σ_i r_i(p)²

The Levenberg-Marquardt algorithm (Levenberg 1944, Marquardt 1963) interpolates between two simpler methods. Gauss-Newton uses the Jacobian J_ij = ∂f(t_i)/∂p_j and solves the normal equations:

(J^TJ) · Δp = J^Tr

which is fast near the optimum but unstable far from it (J^TJ can be near-singular). Steepest descent uses just the gradient and is robust but very slow near the optimum. LM combines them by adding a damping term λ (Marquardt parameter, not the lag time):

(J^TJ + λ · diag(J^TJ)) · Δp = J^Tr

When λ is large, the step approximates steepest descent (small, robust); when λ is small, it approximates Gauss-Newton (large, fast). After each successful step, λ is reduced by a factor of 10; after each failed step (where SSE increases), λ is multiplied by 10. The calculator initializes λ = 10⁻³ and uses central-difference numerical derivatives to compute J. Convergence is declared when the relative change in SSE falls below 10⁻⁹ or after 200 iterations.

Initial guesses are derived from data heuristics: X₀ from the first data point, X_max from the 90th percentile, μ_max from the steepest log-slope across consecutive points, λ from the first time biomass exceeds X₀ by 5%. Poor initial guesses can trap LM at local optima — the auto-select feature mitigates this by trying all five models in parallel.

4 — Goodness-of-fit metrics

Three metrics are reported for each fit, addressing different questions.

4.1 RMSE (root-mean-square error)

RMSE = √(SSE / n)

The typical residual size in the same units as X. Best read alongside the data scale: an RMSE of 0.1 g/L is excellent for X_max = 4 g/L data but poor for X_max = 0.1 g/L data.

4.2 R² (coefficient of determination)

R² = 1 − SSE / SST,  SST = Σ (X_i − X̄)²

The fraction of variance the model explains. R² is unit-independent and roughly comparable across datasets, but it is biased upward as parameter count grows: a 4-parameter model will always reach a higher R² than a 3-parameter model on the same data even when the extra parameter contributes nothing useful.

4.3 AIC (Akaike Information Criterion)

AIC = n · ln(SSE/n) + 2k

where k is the number of parameters. AIC penalizes parameter count: a more complex model only "wins" if its SSE drops by enough to overcome the 2k penalty. Lower is better. AIC is the right metric for choosing between 3-parameter and 4-parameter models — R² alone will spuriously favor the 4-parameter ones.

For small samples (rule of thumb: n / k < 40), use the corrected AICc:

AICc = AIC + 2k(k+1) / (n − k − 1)

which adds an additional small-sample penalty and is reported alongside AIC.

5 — Confidence intervals: Wald approximation

The covariance matrix of the parameter estimates is approximated as:

Cov(p̂) ≈ σ̂² · (J^TJ)⁻¹,  σ̂² = SSE / (n − k)

The standard error of each parameter is the square root of the corresponding diagonal element. The 95% confidence interval is:

p̂_j ± t_{0.975, n−k} · SE(p̂_j)

where t_{0.975, n−k} is the two-sided t-critical value for n − k degrees of freedom. The calculator computes t-critical values via a Wilson-Hilferty approximation. The Wald CI assumes the SSE landscape is locally quadratic around the optimum — a reasonable approximation for batch growth curves with moderate noise, but it can underestimate true uncertainty for parameters that are weakly identified by the data (e.g., λ when no points lie clearly within the lag phase).

6 — Auto-select algorithm

When invoked, the auto-select feature fits all five models in sequence, then ranks the successful fits in two stages: primary by R² descending, tiebreak by AIC ascending. Failed fits (singular Jacobian, non-finite parameters) are listed at the bottom. The user can override the auto-choice by clicking Pick on any other row in the ranking table.

The two-stage ranking handles the common case where two models give nearly identical R² — AIC then picks the simpler one. It does not perfectly mirror AIC-only ranking: occasionally a 4-parameter model will dominate R² by a fraction of a percent while having higher AIC. In those cases the auto-choice is the higher-R² model; for principled model selection on parameter parsimony, use AIC directly.

7 — Number of generations Z

The number of population doublings from inoculum to capacity:

Z = log₂(X_max / X₀)

Z counts the elementary cell-division events the population underwent, regardless of how fast each one was. It is independent of μ_max and λ. Z matters in three contexts:

• Yeast lineage tracking. Petite-mutant accumulation, plasmid loss, and certain epigenetic drifts scale with the cumulative number of doublings, not with elapsed time. A culture that ran for 24 h with Z = 10 carries more genetic baggage than one that ran for 48 h with Z = 5.
• Brewing pitch-rate planning. Successive yeast harvest-and-repitch cycles compound Z; commercial brewers track total Z across generations and discard slurries past a threshold (typically Z > 60–100).
• Population genetic experiments. Mutation rates per generation are calibrated against Z, not chronological time.

8 — Phase durations

Three durations are reported, all derived from the fitted curve rather than from raw data:

• Lag time λ: the explicit fitted parameter (Modified Gompertz, Baranyi-Roberts) or zero (Logistic, Gompertz, Richards). Operationally defined as the intercept where the maximum tangent meets the X₀ level.
• Exponential phase duration: the time interval between λ and the time at which X reaches 95% of X_max. The 95% threshold is conventional; a higher threshold (99%) gives a longer "exponential" duration that bleeds into the deceleration phase.
• Stationary phase duration: the time interval between the end of exponential phase (X = 0.95 X_max) and the last data point. This is a duration of observation, not a duration of stationary biology — if the experiment ended before the cells entered death phase, no death phase is captured.

9 — Yield coefficients and specific rates

When a substrate column is present, the batch-integral yield is:

Y_X/S = (X_max − X₀) / (S₀ − S_final)

This combines biomass produced through both growth and maintenance, so it is always less than the true growth-coupled yield Y_X/S^max. To separate the two, use Tab 2 (Fed-Batch) or Tab 3 (Continuous) which run a Pirt regression across multiple operating points.

When a product column is also present, Y_P/S is computed analogously. The specific rates q_S and q_P are reported as windowed averages over the exponential phase:

q_S = ΔS / (X̄ · Δt),  q_P = ΔP / (X̄ · Δt)

using the first half of the dataset as the window. These are batch-integral approximations and don't replace continuous-culture chemostat measurements for parameter precision.

10 — Metabolism classifier

When yields are computed, the calculator classifies the metabolism into one of four regimes based on Y_X/S and Y_P/S envelopes for S. cerevisiae:

• Aerobic respiratory: Y_X/S ≥ 0.40 g/g, Y_P/S ≤ 0.05 g/g. Glucose is fully oxidized to CO₂ + H₂O via the TCA cycle; ethanol is not produced. Typical of low-glucose chemostat or restricted fed-batch cultivation.
• Aerobic respiro-fermentative: 0.10 ≤ Y_X/S < 0.40 g/g, 0.10 ≤ Y_P/S ≤ 0.40 g/g. Mixed metabolism — partial respiration alongside ethanol overflow under high-glucose aerobic conditions (Crabtree effect).
• Anaerobic fermentative: Y_X/S ≤ 0.15 g/g, Y_P/S ≥ 0.40 g/g. Glucose is converted almost entirely to ethanol + CO₂ via the Embden-Meyerhof-Parnas pathway; biomass yield is low because only 2 ATP per glucose are recovered (vs ~32 ATP in respiration). Typical of brewery and bioethanol fermentations.
• Indeterminate: yields fall outside the canonical envelopes — might reflect mixed conditions, atypical substrates (e.g., xylose for engineered strains), aeration transitions, or under-sampled batches.

The thresholds are operational, not mechanistic — they are based on canonical glucose-cerevisiae behavior and may misclassify other organisms or non-glucose substrates. The classifier is informational; the user can override it via the metabolism dropdown.

11 — When the models break

The five sigmoidal models all assume monotonic growth-and-stationary. Three patterns systematically violate the assumptions:

• Death phase. Post-stationary biomass decline is not captured — including death-phase data in the fit will distort X_max downward and inflate residuals. Truncate to the stationary plateau before fitting.
• Diauxic growth. Two-phase growth (e.g., glucose → ethanol re-utilization) gives a double-S curve that no single model fits well. Use Tab 4 (Diauxic) instead, which detects the phase boundary and fits each phase independently.
• Catastrophic shifts. Sudden μ collapse (oxygen limitation, pH crash, contamination) breaks the smooth-curve assumption. Inspect residuals visually; structured residuals (a curve in the residuals plot) means the model is misspecified for the data, not just noisy.

Reasonable fit quality (R² ≥ 0.95) plus structureless residuals are jointly necessary to trust the parameter values. Either alone is insufficient.

Growth-curve models

• Zwietering, M. H., Jongenburger, I., Rombouts, F. M., & Van 't Riet, K. (1990). Modeling of the bacterial growth curve. Applied and Environmental Microbiology, 56(6), 1875–1881.
• Baranyi, J., & Roberts, T. A. (1994). A dynamic approach to predicting bacterial growth in food. International Journal of Food Microbiology, 23(3–4), 277–294.
• Richards, F. J. (1959). A flexible growth function for empirical use. Journal of Experimental Botany, 10(2), 290–301.
• Gompertz, B. (1825). On the nature of the function expressive of the law of human mortality. Philosophical Transactions of the Royal Society, 115, 513–583.
• Verhulst, P.-F. (1838). Notice sur la loi que la population poursuit dans son accroissement. Correspondance Mathématique et Physique, 10, 113–121. [logistic]

Nonlinear regression

• Levenberg, K. (1944). A method for the solution of certain non-linear problems in least squares. Quarterly of Applied Mathematics, 2(2), 164–168.
• Marquardt, D. W. (1963). An algorithm for least-squares estimation of nonlinear parameters. SIAM Journal, 11(2), 431–441.
• Nocedal, J., & Wright, S. J. (2006). Numerical Optimization (2nd ed.), Chapter 10. Springer.

Industrial yeast kinetics

• Walker, G. M., & Stewart, G. G. (2016). Saccharomyces cerevisiae in the production of fermented beverages. Beverages, 2(4), 30.
• Pirt, S. J. (1975). Principles of Microbe and Cell Cultivation. Blackwell Scientific Publications.
• Bailey, J. E., & Ollis, D. F. (1986). Biochemical Engineering Fundamentals (2nd ed.). McGraw-Hill.
• Doran, P. M. (2013). Bioprocess Engineering Principles (2nd ed.). Academic Press.

Companion FermAxiom tools

• Yeast Growth Rate Calculator — predicts μ under specified glucose / ethanol / temperature / pH using the multiplicative factor decomposition. Use to project the operating envelope after extracting μ_max here.
• Yeast Propagation Simulator — five-tool integrated suite for media design, propagation regime planning, and end-product performance prediction. Y_X/S extracted here feeds into the Medium Composition Calculator.