RESOCONTI /14
P oceedings o he 3 d IAGC
In e na ional Con e ence
Wa e -Rock In e ac ion - 18 & Applied Iso ope Geochemis y - 15
Caglia i, I aly, 16-21 June 2025
edi ed by
F anco F au, Richa d B. Wan y,
Elisabe a Do e, Da io Fancello
UNICAp ess/a eneo
The hi d IAGC In e na ional Con e en-
ce, held in Caglia i (Sa dinia, I aly) in
June 16-21, 2025, was widely a ended
by esea che s om a ound he wo -
ld. This esul ed in he p esen olume
o p oceedings con aining as many as
308 abs ac s, bo h sho and ex en-
ded, dis ibu ed among 22 sessions in
addi ion o he plena y sessions.
The
subjec s co e ed in he abs ac s span
nume ous opics in geochemis y, om
low o high empe a u e, om ozen o
hyd o he mal sys ems, and na u al and
manmade en i onmen s.
The con e ence was hos ed by he De-
pa men o Chemical and Geological
Sciences a he Uni e si y o Caglia i
(UNICA).
UNICAp ess/a eneo
RESCONTI
14
P oceedings o he 3 d IAGC
In e na ional Con e ence
Wa e -Rock In e ac ion - 18 & Applied Iso ope Geochemis y - 15
Caglia i, I aly, 16-21 June 2025
edi ed by
F anco FRAU, Richa d B. WANTY,
Elisabe a DORE, Da io FANCELLO
Caglia i
UNICAp ess
2025
IAGC-3 O ganizing Commi ee
Gio anni B. De Giudici – Sec e a y Gene al
S e ania Da Pelo, F anco F au, Elisabe a Do e, S e ano Nai za, Da io Fancello, Daniela Medas,
Pa izia Onnis, F ancesca Podda, Ricca do Biddau, Elodia Musu
Sil io Fe e o, Lau a Pioli, Richa d B. Wan y, B una B. Ca alho, Mau izio Tes a, Robe o Dessì,
Elisa Sacchi, Lau a Sanna, Pie And ea Ma as
Con ac s:
F au F. ( au @unica.i )
Depa men o Chemical and Geological Sciences, Uni e si y o Caglia i, Caglia i, I aly
Wan y R.B. ( wan [email protected])
Colo ado School o Mines, Golden, CO 80401, USA
Do e E. ([email p o ec ed])
Depa men o Chemical and Geological Sciences, Uni e si y o Caglia i, Caglia i, I aly
Fancello D. (da io. [email protected] )
Depa men o Chemical and Geological Sciences, Uni e si y o Caglia i, Caglia i, I aly
Sezione A eneo
RESOCONTI /14
ISSN 2974-6671
P oceedings o he 3 d IAGC In e na ional Con e ence
Wa e -Rock In e ac ion - 18 & Applied Iso ope Geochemis y - 15
Caglia i, I aly, 16-21 June 2025
edi ed by F anco F au, Richa d B. Wan y, Elisabe a Do e, Da io Fancello
Co e image: De il’s saddle wi h lamingo by Da io Fancello and Emanuele Pucci
© Au ho s and UNICAp ess 2025
CC-BY-SA 4.0
(h ps://c ea i ecommons.o g/licenses/by-sa/4.0/)
Caglia i, UNICAp ess, 2025 (h p://unicap ess.unica.i )
e-ISBN 978-88-3312-187-1
DOI h ps://doi.o g/10.13125/unicap ess.978-88-3312-187-1
P oceedings o he 3 d IAGC In e na ional Con e ence
Wa e Rock In e ac ion-18 & Applied Iso ope Geochemis y-15
243
Towa d g ea e anspa ency: S able iso opes and explainable
Machine Lea ning solu ions o ood aceabili y
Og inc, N.1,*, D ole, J.2, and E imo , T.2
1: Depa men o En i onmen al Sciences, Jože S e an Ins i u e, Ljubljana, Slo enia
2: Depa men o Compu e Sys ems, Jože S e an Ins i u e, Ljubljana, Slo enia
*Co esponding au ho : ni es.og [email protected]
ABSTRACT
The g owing demand o au hen ic and high-quali y ood has d i en he need o
ad anced aceabili y sys ems o e i y geog aphical o igin and p e en aud. This
s udy in eg a es s able iso ope and elemen al analysis wi h machine lea ning (ML) o
enhance ood au hen ica ion. By employing ML echniques o da a analysis, ea u e
selec ion, and classifica ion, we imp o e accu acy and p o ide explainable insigh s
in o ood p o enance. A case s udy on saff on demons a es he me hod’s effec i e-
ness, achie ing up o 90% accu acy in o igin classifica ion. This app oach offe s a scal-
able and adap able solu ion o a ious ood p oduc s, s eng hening ood sa e y,
anspa ency, and egula o y compliance.
INTRODUCTION
The connec ion be ween ood and e i o y has g adually been los o e ime due
o ad ances in p oduc ion, anspo a ion, and exposu e o globaliza ion (Luykx and
an Ru h, 2008). In esponse, consume demand o au hen ic, high-quali y, and sus-
ainably p oduced ood is ising. T adi ional cul i a ion and p ocessing me hods a e
alued o hei au hen ici y, sa e y, en i onmen al benefi s, and quali y, o en com-
manding p emium p ices.
Among he exploi able echniques, s able iso ope and elemen al finge p in ing a e
leading he way in es ablishing au hen ici y and geog aphical o igin o ood p oduc s
(Danezis e al., 2016). The basis o he s able iso ope app oach lies in he ans e o he
iso opic signals (iso opic finge p in ) o bio-elemen s (H, C, N, O, S) p esen in local
eed and wa e o animal and plan issue, p ocesses ha a e gene ally well unde -
s ood. The e ifica ion o egional o igin can be e en mo e effec i e when s able iso-
opes a e combined wi h elemen al composi ion since a plan 's elemen al p ofile is
ela ed o he soil composi ion o he loca ion whe e he plan g ows; consequen ly,
bioa ailable nu ien s can p o ide di ec in o ma ion abou an ag icul u al p oduc s'
geog aphical o igin (D i elos and Geo giou, 2012).
Howe e , s able iso opes and elemen al finge p in ing alone a e insufficien . A
comp ehensi e e e ence da ase o au hen ica ed p oduc s is essen ial bu challeng-
ing o es ablish due o high ime and cos demands. This da ase mus include a wide
ange o samples ep esen ing di e se geog aphical, seasonal, die a y, and p oduc ion
P oceedings o he 3 d IAGC In e na ional Con e ence
Wa e Rock In e ac ion-18 & Applied Iso ope Geochemis y-15
244
condi ions (Kelly e al., 2005). Au hen ici y is assessed by compa ing alues om com-
me cial samples agains da ase -de i ed limi s, using a s a is ical model o de e mine
he bes fi .
The p esen a ion will desc ibe a comp ehensi e au hen ici y and aceabili y sys-
em ha in eg a es s able iso ope analysis, elemen al composi ion, da abases, s a is i-
cal e alua ions, and eme ging AI-d i en echnologies, pa icula ly machine lea ning
(ML) applica ions. Special emphasis will be placed on how ML enhances da a p o-
cessing, pa e n ecogni ion, and aud de ec ion in ood au hen ica ion. In he final
sec ion, a eal-wo ld implemen a ion o hese echnologies in a aceabili y sys em will
be p esen ed, demons a ing hei p ac ical impac .
METHODS
The analysis o s able iso ope a ios and elemen al composi ion was conduc ed us-
ing p ecise and well-es ablished echniques. The wo main analy ical me hods used
we e Iso ope Ra io Mass Spec ome y (IRMS) and Induc i ely Coupled Plasma Mass
Spec ome y (ICP-MS) o elemen al p ofiling.
DATABASE
A comp ehensi e da abase and da abase managemen sys em (DBMS) called
IsoFoodT ack (h p://iso ood ack.ijs.si) was de eloped (Te o e al., 2025). This sys em
p o ides ex ensi e da a on he s able iso opes o ligh elemen s and he elemen al
composi ion o au hen ic samples om a ious ood commodi ies such as oils, milk
and dai y p oduc s, mea , spices, uffles, sea ood and ege ables. Fu he mo e,
IsoFoodT ack is designed o be in e ope able, allowing connec ion wi h o he da a-
bases o cen alized eposi o ies and ep esen s a significan ad ancemen o e adi-
ional ood da abases by p io i izing bo h accessibili y and s anda diza ion. Addi ion-
ally, he da abase in eg a es s anda dized me ada a p o ocols and ha monized da a
en y o ma s, which s eamline c oss-s udy compa isons and enhance ep oducibil-
i y.
MACHINE LEARNING ENHANCED DATA PROCESSING
Al hough classical s a is ical app oaches ha e his o ically been used in ood ace-
abili y esea ch, machine lea ning (ML) me hods a e now inc easingly employed and
gaining popula i y. In mos cases, hese s udies use iso opic measu emen s o ood
samples o p edic o classi y hei geog aphic o igin, employing classifica ion models
such as Random Fo es , XGBoos , o Neu al Ne wo ks and compa ing hei pe o -
mance me ics. Despi e an abundance o published li e a u e and g owing p essu e
om e iewe s o include hese analyses, many s udies simply apply ML models
wi hou adhe ing o a igo ous expe imen al design o p ope benchma king, high-
ligh ing he need o s anda dized me hods in ML-based ood aceabili y esea ch.
P oceedings o he 3 d IAGC In e na ional Con e ence
Wa e Rock In e ac ion-18 & Applied Iso ope Geochemis y-15
245
To add ess he a o emen ioned challenges and p o ide deepe insigh s in o ood
aceabili y, we p opose a no el ML-based amewo k consis ing o he ollowing
s eps: (1) selec ing ep esen a i e aining da a, (2) pe o ming ea u e selec ion o
iden i y measu emen s ha imp o e model pe o mance wi hou o e fi ing, (3) ain-
ing he model, (4) e alua ing he model using di e se ain– es spli s o ensu e obus
esul s, and (5) p o iding sample-specific explana ions h ough ea u e impo ance.
Selec ing ep esen a i e aining da a may elimina e he need o use all a ailable
samples. Ins ead, we epea edly selec aining and es subse s using unsupe ised
lea ning o g aph-based me hods o ensu e an app oxima ely uni o m dis ibu ion o
samples in he ea u e space. Nex , we apply ea u e selec ion echniques (e.g., co e-
la ion analysis) o iden i y he measu emen s ha yield imp o ed classifica ion mod-
els wi hou o e fi ing. The model is hen ained on each ain– es spli using he
selec ed ea u es. To enhance explainabili y, we pe o m SHAP (SHapley Addi i e ex-
Plana ions) analysis on each sample in he es se , he eby e ealing he in e ac ions
and ela i e impo ance o ea u es ha con ibu e o o igin classifica ion o each
sample sepa a ely. This app oach elucida es he ea u es mos c i ical o accu a e p e-
dic ions and highligh s pa e ns leading o misclassifica ions. By agg ega ing explana-
ions ac oss all es spli s, we iden i y obus insigh s ha a e unlikely o occu by
chance. Ul ima ely, hese esul s gene a e a “ ood o igin oo p in ”, indica ing he
mos influen ial ea u es and hei in e ac ions o each o igin.
PRACTICAL APPLICATION
The example o p ac ical applica ion includes he saff on da a sou ced om he
main saff on-p oducing coun ies: I an, I aly, Spain, Mo occo and G eece. Specifically,
we ained obus classifie s capable o de ec ing he geog aphical o igin o he saff on
samples. We es ed diffe en scena ios using only iso opic a iables, ace elemen s
and a combina ion o bo h. By epea ing he lea ning p ocess fi e imes o assess he
obus ness o he model ac oss all scena ios, we achie ed a e age accu acy o a ound
90% ac oss he spli s.
Addi ionally, h ough explainable pos -hoc analysis (SHAP), we iden ified which
a iables a e impo an o each coun y o ob ain eliable o igin diffe en ia ion. The
model's accu acy was 80% when only s able iso ope analysis was included in he e al-
ua ion. The mos eliable pa ame e s o Mo occo and I aly a e p esen ed in Figu e 1.
Fo bo h coun ies, δ18O alues con ibu e he mos o hei sepa a ion, ollowed by
δ13C in I aly and δ34S in Mo occo. This analysis has been done o all he coun ies o
o igin, bu only o he co ec ly p edic ed samples, since hey gi e insigh in o which
a ibu es con ibu e o he co ec p edic ion.
P oceedings o he 3 d IAGC In e na ional Con e ence
Wa e Rock In e ac ion-18 & Applied Iso ope Geochemis y-15
246
Figu e 1: Decision plo analysis o Mo occo and I aly showing how specific iso opic a ios a e c i i-
cal in dis inguishing saff on o igin.
When bo h iso opic and elemen al da a a e conside ed, he sepa a ion imp o es
sligh ly, eaching an accu acy o 89.8%, wi h he mos disc imina i e pa ame e s being
he elemen al a iables. Al hough his app oach has been used o saff on as an exam-
ple, i is eadily adap able o o he aw ma e ials.
CONCLUSIONS
This s udy demons a es he effec i eness o in eg a ing s able iso ope and ele-
men al analysis wi h machine lea ning (ML) o ood aceabili y. The me hod is
adap able o a ious ood commodi ies, offe ing a scalable solu ion o s eng hening
ood quali y con ol, au hen ici y e ifica ion, and aud p e en ion in he global ood
supply chain.
ACKNOWLEDGEMENTS
We acknowledge he financial assis ance o he Ho izon Eu ope p ojec s FishEUT-
us (G an ag eemen no. 101060712) and PROMEDLIFE p ojec PRIMA p og amme
suppo ed by he Eu opean Union (G an ag eemen No. 2132).
REFERENCES
Danezis, G.P., Tsagka is, A.S., Camin, F., B usic, V., and Geo giou, C.A., 2016. Food au hen ica ion:
echniques, ends & eme ging app oaches. T ends in Analy ical Chemis y, .85, pp. 123–132.
D i elos, S.A., and Geo giou, C.A., 2012. Mul i-elemen and mul i-iso ope- a io analysis o de e mine
he geog aphical o igin o oods in he Eu opean Union. T ends in Analy ical Chemis y, .40, pp. 38–
51.
Kelly, S., Hea on, K., and Hoogewe ff, J., 2005. T acing he geog aphical o igin o ood: he applica ion
o mul i-elemen and mul i-iso ope analysis. T ends in Food Science and Technology, .16, pp. 555–
567.
Luykx, D.M.A.M., and an Ru h, S.M., 2008. An o e iew o analy ical me hods o de e mining he
geog aphical o igin o ood p oduc s. Food Chemis y, .107(2), pp. 897-911.
Te o, C., Modic, R., Og inc, M., Simčič, A., D ole, J., E imo , T., Ko oušić Seljak, B., and Og inc, N.,
2025. IsoFoodT ack: a comp ehensi e da abase and managemen sys em based on s able iso ope
a io analysis o comba ing ood aud. F on ie s in Nu i ion, .12, 1516521.
